0 Votes +0 Votes -

76 vues168 pagesMean-Value Modelling and Robust Control of the Airpath of a Turbocharged Diesel Engine

May 31, 2017

© © All Rights Reserved

PDF, TXT ou lisez en ligne sur Scribd

Mean-Value Modelling and Robust Control of the Airpath of a Turbocharged Diesel Engine

© All Rights Reserved

76 vues

Mean-Value Modelling and Robust Control of the Airpath of a Turbocharged Diesel Engine

© All Rights Reserved

Vous êtes sur la page 1sur 168

Merten Jung

Sidney Sussex College

Department of Engineering

University of Cambridge

February 2003

ii

In memory of my grandparents,

Hildegard and Josef Jung

Hopkinson Lab

Department of Engineering

Trumpington Street

Cambridge CB2 1PZ

United Kingdom

Fax: +44 (0)1223 332662

E-mail: Jung@tu-harburg.de

Abstract

This thesis investigates mean-value modelling and robust control of the airpath of a diesel engine

equipped with variable geometry turbocharger (VGT) and exhaust gas recirculation (EGR). Both

the EGR and the VGT are driven by the exhaust gas and render the control problem inherently

multivariable. In order to allow for model-based control design, a mean-value model of the airpath

of the engine is derived with a focus on the parameterisation of the turbocharger. The effect of heat

transfer via the turbocharger housing on the efficiencies derived from temperature measurements

is identified as being very significant at low speed and load points regularly met on emission drive

cycles. A physics-based parameterisation for the turbine efficiency map, which is the most difficult to

model, is suggested. Based on experimentally obtained frequency responses, the parameters which

are most uncertain in the model, i. e. compressor and turbine efficiency as well as the effective area

of the EGR valve, are identified to have the biggest effect on the frequency responses. Different

parameterisations of these uncertainties are then used for extended H loopshaping design at a

fixed engine operating point. Applying analysis tools, it turns out that the application tailored

uncertainties yield a better controller performance, which is confirmed by experimental data.

In order to extend the controller operating regime, the nonlinear model is simplified and converted

to linear parameter-varying (LPV) form. A robustly gain scheduled LPV controller is synthesised

for this model using a gridding approach for the intake manifold pressure as scheduling variable.

The designed controller is implemented on the engine in real-time. The experimental results are

very promising and indicate that the quasi-LPV model captures the significant nonlinearities and

dynamics of the plant.

Keywords: Diesel Engine, Airpath, Turbocharger, Mean-Value Models, Robust Control, H Loop-

shaping, synthesis, Linear Parameter Varying, Gain Scheduling, Experimental Engine Study

iii

iv Abstract

Preface

Three years of Cambridge life come to an end and it remains to take the opportunity to acknowledge

the people who made this time so enjoyable. Firstly, I would like to thank my supervisor, Prof. Keith

Glover, and my adviser, Prof. Nick Collings, for their guidance, support, and especially for estab-

lishing the perfect environment for any student who finds an interest in engine control (ranging from

not for the faint-hearted robust control theory to greasy hands after a day of fiddling around with

engines).

The research conducted in this thesis is the outcome of a cooperation with the Ford Motor

Company who funded the diesel engine facilities and who also contributed towards my maintenance

allowance. Particular thanks go to Dave Ketcher, Mike Watts, Robert Marshall, and Trevor Taylor.

Urs Christen from Ford Forschungszentrum Aachen deserves special mention for so many very

useful discussions and suggestions which have had a significant impact on the final outcome of this

thesis. Technical support and permission to publish turbocharger data from Honeywell-Garrett is

also acknowledged.

I am very grateful to Katie Vantine for introducing me to the engine test facilities. Particular

thanks go to Dr. Richard Ford who has actively supported my research from the very beginning and

who has taught me many skills on how to run the test cell. I admire his patience and considering

even most annoying practical problems as an intriguing opportunity to learn more. The day-to-day

operation of the engine test facilities would not be possible without the skilled team of technicians

at the University of Cambridge. Particular thanks go to Dave Gautrey, Trevor Parsons, John Harvey,

and Michael Underwood.

My stay in Cambridge would not have been possible without the generous funding of the European

Commission through the program Training and Mobility of Researchers Research Networks

and through project System Identification (FMRX CT98 0206). I would also like to acknowledge

contacts with the participants in the European Research Network System Identification (ERNSI).

My co-workers and the faculty in the control and engine groups certainly deserve a mention

since they have significantly contributed not only to my work but also to my social welfare. Special

thanks to Dr. Kelvin Halsey and Christos Papageorgiou who were always willing to help me.

My personal view is that Cambridge life is not just about work, but about combining academic

excellence with a pleasant social life. The latter would not have been possible without good friends.

Here in Cambridge, special thanks go to Umut Gen for many discussions about the sense of life and

relentless clubbing afterwards, to Cory Sutela for going out for single beers, and to Torsten Riotte

v

vi Preface

for convincing me that one has not really been at Cambridge University without having rowed for

at least one term. Back home in Germany, I would like to thank all the friends in Wilhelmshaven

and hope that we will always find opportunities to get together independent of where in the world

we will end up.

Finally but most importantly, I would like to say thank you to my grandparents and my mum

for continuous support and love throughout my life, and to Annika Uhlendorf for making the good

times much more fun and the bad times much more bearable.

Merten Jung

February 2003

As required by the University Statute, I hereby declare that this dissertation is the result of my

own work and includes nothing which is the outcome of work done in collaboration except where

specifically indicated in the text.

This dissertation contains no more than 65,000 words and 150 figures.

Contents

List of Figures xi

List of Tables xv

Notation xvii

Acronyms xix

1 Introduction 1

1.1 Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.1 Diesel Engines . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.1.2 Modern Control Systems . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

1.2 Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.1 Formation of Nitrogen Oxides (NOx ) and Particulate Matter (PM) . . . . . . . . . 12

2.2 Exhaust Gas Recirculation (EGR) . . . . . . . . . . . . . . . . . . . . . . . . . . 14

2.3 NOx Traps (Adsorbers) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

2.3.1 Operating Principle . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.3.2 Trap Regeneration by Rich Combustion . . . . . . . . . . . . . . . . . . . 18

2.3.3 Effects of Sulphur and Desulphation Strategies . . . . . . . . . . . . . . . 19

2.4 Alternative DeNOx Solutions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.4.1 DeNOx Catalysts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.4.2 Selective Catalytic Reduction (SCR) . . . . . . . . . . . . . . . . . . . . . 21

2.4.3 Non-thermal Plasma Technology . . . . . . . . . . . . . . . . . . . . . . . 23

2.5 Particulate Matter . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.5.1 Nature of Diesel Particulates . . . . . . . . . . . . . . . . . . . . . . . . . 23

2.5.2 Particulate Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

2.6 Integrated DeNOx and PM solutions . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.6.1 Filters and SCR/NOx Traps . . . . . . . . . . . . . . . . . . . . . . . . . 27

2.6.2 Fuel Enhancement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

2.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

vii

viii Contents

3.1 Motivation and Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

3.2 Mean-Value Engine Modelling . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.2.1 Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.2.2 Model Conversion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32

3.2.3 Turbocharger . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33

3.2.4 Other Subsystems . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

3.2.5 Overall Diesel Engine Model . . . . . . . . . . . . . . . . . . . . . . . . . 36

3.3 Experimental Turbocharger Maps . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.3.1 Compressor Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37

3.3.2 Compressor Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40

3.3.3 Turbine Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.3.4 Turbine Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 42

3.4 Turbocharger Parameterisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.4.1 Compressor Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.4.2 Compressor Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

3.4.3 Turbine Flow . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.4.4 Turbine Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44

3.5 Experimental Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47

4.1 The EGR-VGT Control Problem . . . . . . . . . . . . . . . . . . . . . . . . . . . 51

4.1.1 Effects of Linearisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52

4.1.2 Plant Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4.2 Frequency Domain Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56

4.2.1 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57

4.2.2 Actuator Position Control Loops . . . . . . . . . . . . . . . . . . . . . . . 58

4.2.3 Plant Dynamics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59

4.3 Model Uncertainty Investigation . . . . . . . . . . . . . . . . . . . . . . . . . . . 61

5.1 H Loopshaping Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63

5.1.1 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.1.2 Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

5.1.3 Controller Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67

5.2 Two Degrees of Freedom Controller for Coprime Uncertainty . . . . . . . . . . . . 68

5.2.1 Control Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69

5.2.2 Control Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

5.2.3 Controller Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72

5.3 Two Degrees of Freedom Controller for Additive Uncertainty . . . . . . . . . . . . 73

5.3.1 Uncertainty Parameterisation . . . . . . . . . . . . . . . . . . . . . . . . . 73

ix

5.3.3 Controller Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 77

5.4 Alternative Uncertainty Parameterisations . . . . . . . . . . . . . . . . . . . . . . 78

5.4.1 Parametric State-Space Uncertainty . . . . . . . . . . . . . . . . . . . . . 79

5.4.2 Pulling Out the Deltas Approach . . . . . . . . . . . . . . . . . . . . . . 79

5.5 Controller Comparison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

5.5.1 Mixed Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81

5.5.2 Load Step Responses . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83

5.6 Mixed Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85

6.1 Simplified Nonlinear Model . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 90

6.2 Conversion to a Quasi-LPV Model . . . . . . . . . . . . . . . . . . . . . . . . . . 93

6.3 Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95

7.1 Theoretical Background . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99

7.2 LPV Loopshaping Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101

7.2.1 Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7.2.2 Choice of Weights . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103

7.2.3 Parameter Dependence and Gridding . . . . . . . . . . . . . . . . . . . . 105

7.2.4 Controller Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

7.2.5 Controller Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

7.2.6 Implementation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106

7.2.7 Controller Performance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

7.3 Controller Extensions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7.3.1 Feedforward . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7.3.2 Disturbance Rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110

7.4 Experimental Results from NEDC . . . . . . . . . . . . . . . . . . . . . . . . . . 112

7.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116

8 Conclusions 119

8.1 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

8.2 Further Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120

A.1 Fundamentals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

A.1.1 State-space Representations . . . . . . . . . . . . . . . . . . . . . . . . . 123

A.1.2 Function Spaces . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123

A.1.3 Coprime Factorisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124

A.1.4 Internal Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125

x Contents

A.1.6 The General Framework . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

A.2 Controller Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

A.2.1 H Optimal Control . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 129

A.2.2 H Loopshaping . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130

A.3 Advanced Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

A.3.1 Structured Singular Value . . . . . . . . . . . . . . . . . . . . . . . . . . 132

A.3.2 Gap Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 132

B.1 Hardware . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

B.1.1 Dynamometer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

B.1.2 Engine . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

B.2 Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135

B.3 Hardware/Software Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136

Bibliography 143

List of Figures

1.2 Diesel engine configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

1.3 Standard feedback configuration . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.2 Quasi-steady flame . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

2.3 Operating principle of NOx traps . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2.4 Typical NOx trap temperature window . . . . . . . . . . . . . . . . . . . . . . . . 17

2.5 Soot generating region on the -T-map . . . . . . . . . . . . . . . . . . . . . . . . 19

2.6 Particle size distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

3.2 Compressor flow map and drivecycle data . . . . . . . . . . . . . . . . . . . . . . 31

3.3 Simulink block diagram of the diesel engine model . . . . . . . . . . . . . . . . . 38

3.4 Compressor Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 39

3.5 Turbine Maps . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41

3.6 Turbine efficiency due to heat transfer . . . . . . . . . . . . . . . . . . . . . . . . 45

3.7 Turbine aerodynamic efficiency map . . . . . . . . . . . . . . . . . . . . . . . . . 46

3.8 Cooling efficiency of the turbocharger . . . . . . . . . . . . . . . . . . . . . . . . 47

3.9 Validation results for NEDC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48

3.10 Validation results for MAP and MAF . . . . . . . . . . . . . . . . . . . . . . . . . 49

4.2 MAF and MAP step responses . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54

4.3 MAF and MAP step responses for closed EGR . . . . . . . . . . . . . . . . . . . 57

4.4 Frequency domain validation methodology . . . . . . . . . . . . . . . . . . . . . 58

4.5 Identified frequency responses of the actuator position control loops . . . . . . . . 59

4.6 Frequency response of the scaled plant . . . . . . . . . . . . . . . . . . . . . . . . 60

4.7 Frequency response of the scaled plant with varied parameters . . . . . . . . . . . 62

5.2 Comparison of scaled and weighted plant . . . . . . . . . . . . . . . . . . . . . . 66

xi

xii List of Figures

5.4 Frequency response of the closed loop for the H controller . . . . . . . . . . . . 67

5.5 Validation of the H loopshaping controller . . . . . . . . . . . . . . . . . . . . . 68

5.6 Two degrees of freedom H loopshaping configuration [SP96]. . . . . . . . . . . 69

5.7 Robust performance vs. robust stability . . . . . . . . . . . . . . . . . . . . . . . 70

5.8 Implementation setup for the 2DOF controller . . . . . . . . . . . . . . . . . . . . 71

5.9 Frequency response of the closed loop for the 2DOF controller . . . . . . . . . . . 72

5.10 Validation of the coprime uncertainty controller . . . . . . . . . . . . . . . . . . . 73

5.11 Control setup with additive uncertainty . . . . . . . . . . . . . . . . . . . . . . . . 74

5.12 Frequency response of the shaped plant with varied parameters and bounds . . . . 75

5.13 Performance comparison of CUC and AUC controller . . . . . . . . . . . . . . . . 77

5.14 Frequency response of the closed loop for the additive uncertainty controller . . . . 77

5.15 Validation of the additive uncertainty controller . . . . . . . . . . . . . . . . . . . 78

5.16 Frequency response for state-space uncertainty . . . . . . . . . . . . . . . . . . . 80

5.17 Frequency response for pulled out deltas . . . . . . . . . . . . . . . . . . . . . . 81

5.18 Setup for analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 82

5.19 analysis of CUC and AUC controller . . . . . . . . . . . . . . . . . . . . . . . . 82

5.20 Comparison of ECU to H , CUC, and AUC controller . . . . . . . . . . . . . . . 83

5.21 analysis of SSUC controller . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84

5.22 Frequency response of the closed loop for the state-space uncertainty controller . . 85

5.23 Validation of the state-space uncertainty controller . . . . . . . . . . . . . . . . . . 86

6.1 Static nonlinearity of the EGR valve and flow function approximation . . . . . . . 91

6.2 Turbine flow map for the quasi-LPV model . . . . . . . . . . . . . . . . . . . . . 92

6.3 Validation results for the quasi-LPV model on the NEDC . . . . . . . . . . . . . . 96

6.4 MAF and MAP step responses of the quasi-LPV model . . . . . . . . . . . . . . . 97

7.1 Transient intake and exhaust manifold pressure measurements from NEDC . . . . 102

7.2 Frequency responses of the scaled quasi-LPV model . . . . . . . . . . . . . . . . 104

7.3 Parameter dependence of the A matrix . . . . . . . . . . . . . . . . . . . . . . . . 105

7.4 Frequency response of the LPV controller . . . . . . . . . . . . . . . . . . . . . . 107

7.5 Frequency response of the closed loop . . . . . . . . . . . . . . . . . . . . . . . . 107

7.6 Validation of the LPV loopshaping controller . . . . . . . . . . . . . . . . . . . . 108

7.7 Setup for feedforward . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109

7.8 Modified references with feedforward . . . . . . . . . . . . . . . . . . . . . . . . 110

7.9 Setup for disturbance rejection . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111

7.10 Frequency response for disturbance rejection . . . . . . . . . . . . . . . . . . . . 112

7.11 Response to disturbance steps . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

7.12 LPV performance on extra-urban part of NEDC . . . . . . . . . . . . . . . . . . . 114

7.13 LPV performance on urban part of NEDC . . . . . . . . . . . . . . . . . . . . . . 115

xiii

A.3 Small gain theorem . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

A.4 General framework . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

A.5 Linear fractional transformation framework . . . . . . . . . . . . . . . . . . . . . 128

A.6 General framework for controller synthesis . . . . . . . . . . . . . . . . . . . . . 129

A.7 Framework for controller synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . 130

A.8 The loopshaping design procedure . . . . . . . . . . . . . . . . . . . . . . . . . . 131

xiv List of Figures

List of Tables

7.2 Optimal LPV stability margins for different parameterisations . . . . . . . . . . . 106

B.2 Overview of sensors in engine test cell . . . . . . . . . . . . . . . . . . . . . . . . 136

xv

xvi List of Tables

Notation

Engine

A m2 area

cp J/kg/K specific heat at constant pressure

cu - blade speed ratio

cv J/kg/K specific heat at constant volume

E J energy

h J/k specific enthalpy

J kgm2 inertia

l m length

m kg mass in the manifold

N rpm engine speed

Nt rpm turbocharger shaft speed

p Pa pressure

P W power

Qw J heat transfer

R J/kg/K gas constant

t s time

T K temperature

T Nm torque

u J/kg specific internal energy

V dm3 volume

W J work transfer

Wab kg/s mass flow rate from a into b (time-based)

wab kg/st mass flow rate from a into b (event-based)

x - (normalized) actuator position

- efficiency

- actual AFR normalised by stoichiometric AFR

- short for ( 1)/

kg/m3 density

st crankshaft position

e st/s engine speed (event-based)

xvii

xviii Notation

Subscripts applied to h, m, p, P, T, V, W,

a ambient

b brake

c compressor

d displacement

dr driveline

e engine block

f friction

i intake manifold

ind indicated

l load

r exhaust gas recirculation

t turbine

v volumetric

x exhaust manifold

Control

field of complex numbers

field of real numbers

j imaginary axis

belong to

:= defined as

In n n identity matrix

largest singular value

L2

square integrable functions on j

H2 subspace of L2 with functions analytic in Re(s) > 0

L functions bounded on j

H subspace of L with functions analytic in Re(s) > 0

prefix R real rational, e.g. RH

K controller

M system interconnection matrix for analysis

P (generalized) plant

Tzw transfer function from w to z

u system inputs

W weighting matrix

x system states

y system outputs

uncertainty

structured singular value

singular value

Acronyms

Engine

AFR Air to Fuel Ratio

ASI After Start of Injection

ATDC After Top Dead Centre

BMEP Break Mean Effective Pressure

BPT Balance Point Temperature

CI Compression Ignition

CRT Continuously Regenerating Trap

DI Direct Injection

DPF Diesel Particulate Filter

ECU Engine Control Unit

EGR Exhaust Gas Recirculation

EVRV Electronic Vacuum Regulator Valve

FAR Fuel to Air Ratio

GDI Gasoline Direct Injection

HDD Heavy-Duty Diesel

ICE Internal Combustion Engine

IDI Indirect Injection

LDD Light-Duty Diesel

MAF Mass Air Flow

MAP Manifold Absolute Pressure

NEDC New European Drive Cycle

OP Operating Point

PAH Polycyclic Aromatic Hydrocarbons

PM Particulate Matter

SCR Selective Catalytic Reduction

SI Spark Ignition

SOF Soluble Organice Fraction

SVR Swept Volume Ratio

VNT Variable Nozzle Turbine

VGT Variable Geometry Turbocharger

xix

xx Acronyms

Control

2DOF Two Degrees of Freedom

AUC Additive Uncertainty Controller

CAN Controller Area Network

LFT Linear Fractional Transformation

LMI Linear Matrix Inequality

LPV Linear Parameter-Varying

MIMO Multiple-Input Multiple-Output

PID Proportional, Integral, Derivative

SISO Single-Input Single-Output

1

Introduction

In todays society, mobility plays a major role. On the one hand, it enables us to reach ever remote

places all over the world in less and less time, on the other hand, mobility causes environmental

pollution and is associated with safety risks. The dominant means of transportation in our every-day

life is certainly the passenger cars such that it is hard to imagine life without them. Hence, the

optimisation of passenger cars with respect to emissions, fuel consumption, driving performance,

and safety becomes more and more important.

Decisive progress in attaining these more or less conflicting design objectives would not have

been possible without the development of electronic engine control systems over the past thirty years.

However, the requirements imposed on modern passenger cars have been raised at the same speed:

Firstly, the emission levels, which are already severely regulated by legislature in key markets such

as the US, Europe, and Japan, will become even tighter in the near future (cf. Fig. 2.1). This poses

the automotive companies with a big challenge, especially in the diesel engine sector. The increasing

number of publications in this area, e. g. at the SAE World Congress, reflects this situation. The

present dissertation is also motivated by this trend and aims at providing improved control over the

airpath of a turbocharged diesel engine to allow for better aftertreatment. Secondly, with rising fuel

prices, especially in Europe, the customers demand for fuel-saving cars imposes another requirement

on modern passenger cars which conflicts with some of the aftertreatment methodologies used to

achieve lower emission levels. At the same time, cost, safety, and performance can never be neglected.

Nevertheless, most of the control systems in production type cars are still based on engine map-

ping and classical theory. This is due to practical problems such as cost considerations, calibratability,

and limited on-board computing power. However, the aforementioned demands on modern passen-

ger cars increase the opportunities for modern control methodologies such as H robust control,

which has already been successfully applied in aerospace industry for many years. This dissertation

attempts to show the benefit of such advanced tools for the automotive industry, which is, apart from

improved controller performance, a potential reduction in calibration effort. The latter is crucial in

near future considering the ever shorter development cycles, currently around 24 months for a whole

new passenger car platform.

While new concepts such as fuel cells, hybrid powertrains, lean-burn gasoline direct injection

(GDI), and homogeneous charge compression ignition (HCCI) engines are getting serious attention,

the vast majority of passenger cars today are equipped with either stoichiometric spark-ignition (SI)

gasoline or lean-burn compression-ignition (CI) diesel engines. This dissertation focuses on diesel

1

2 Chapter 1. Introduction

engines, which have made enormous progress over the past decades from a rather noisy and sluggish

to a high-performance engine propelling even luxury cars. This has become possible not least due

to high performance control systems in conjunction with state-of-the-art high-pressure injection

systems such as common rail and unit injector.

The next section will present some of the theoretical background on which this dissertation is

based. An introduction to CI internal combustion engines is given in Section 1.1.1. For a thorough

overview of internal combustion engines, the reader is referred to [Hey88]. Section 1.1.2 introduces

the reader to control engineering and gives some motivation for its use in automotive applications.

An outline of this dissertation is given in Section 1.2.

1.1 Background

Engine control combines amongst others the fields of thermo/fluids and control engineering, two

subjects which are usually studied independently. The general nature of control engineering provides

analysis and design tools that can be applied to a great variety of applications even in non-technical

areas such as finance. The implementation of complex control systems in automotive applications

has become feasible with the introduction of microprocessor based engine management in the 1970s.

Ever since, the impact of control engineering on the automotive industry has been steadily growing.

As a consequence of this process, a collaboration (the so-called Engine Control Group) has been

established at Cambridge University which provides expertise from both fields, thermo/fluids as well

as control.

As opposed to the homogeneous charge spark-ignition (SI) engine, the diesel engine combusts a

compression-ignition (CI) inhomogeneous charge. Our understanding of the diesel combustion

process has evolved over the years and significant progress has been made with optical techniques

such as laser imaging. Flynn et al. combined laser techniques with chemical kinetic models to

postulate a new description of the combustion process in direct injection CI engines [FDH+99].

Figure 1.1 shows a sequence of idealised schematics of the diesel combustion during injection

for different crank angle degrees after the start of injection (ASI). The cold fuel jet penetrates the

combustion chamber and entrains hot air in the cylinder causing the liquid fuel to warm up and to

vapourise. A sheath of vapourised air/fuel mixture (with relative air/fuel ratio 0.25 - 0.5, i. e. four

to two times richer than stoichiometric) encloses the growing fuel jet until it reaches a temperature

of around 750 K at which point the fuel begins to react (5 ASI). These reactions and the additional

hot air entrained into the jet raise its temperature to approximately 825 K increasing the rate of

combustion. As these partial oxidations take place, reaction products are formed, pushed aside by

and then re-entrained into the penetrating jet (heating it up to 1600 - 1700 K). At around 6.5 ASI,

a diffusion flame sheath forms around the cloud of partial oxidation products where temperatures

reach 2700 K. After 10 ASI, the process reaches a quasi-steady state as depicted in the last frame

of Figure 1.1. Eventually, the air entrained during the premixed burn is entirely consumed and the

1.1. Background 3

mixing-controlled combustion in a quasi-steady flame continues until the end of injection, and after.

This model has been found helpful in understanding NOx and particle formulation, and the nature

of the heat release.

Note that the model assumes a continuous single input of fuel which is not valid for multiple

injections realisable with common rail injection systems such as the one in the engine under in-

vestigation in this dissertation. However, the model gives a good qualitative understanding of the

diesel combustion process in general and the formation of pollutants as described in Section 2.1 in

particular.

Compression-ignited diesel engines have a number of advantages over spark-ignited gasoline

engines due to the nature of the combustion process described above. In fact, the direct injection

diesel engine is the most efficient internal combustion engine available. This is due to the following

reasons [Hey88]:

CI engines do not need to throttle the intake air since they are operated with an inhomogeneous

mixture so that the load can be controlled by the amount of fuel injected into the cylinder.

Hence, CI engines avoid pumping losses at part load which results in a higher mechanical

efficiency.

In CI engines, the charge inducted into the cylinder does not contain any fuel which could

possibly ignite unintentionally (knock). Therefore, higher compression ratios are realisable

which results in higher fuel conversion efficiency, since more work is extracted during the

expansion stroke.

4 Chapter 1. Introduction

The CI engine operates always lean (except if short rich excursions are needed for the feedgas

to the aftertreatment devices as will be explained in Section 2) so that the ratio of specific heats

( =c p /cv ) is higher over the expansion process which also results in a higher fuel conversion

efficiency.

CI engines have a higher volumetric efficiency since there are no flow losses at throttle (and

carburetor), no intake manifold heating, no fuel vapour (since the fresh charge does not contain

any fuel), and finally, CI engines have a lower residual gas fraction.

Without the knock problems of gasoline engines, the cylinders of a diesel engine can be made

larger which leads to lower heat losses for engines of comparable power. This is due to the

fact that engine power increases with displacement volume (d3 ), but heat losses with surface

area (d2 ). Thus, big diesel engines (e. g. for marine propulsion) can achieve brake efficiencies

around 50 %.

These effects can add up to as much as 20 % more efficiency of the CI engine with respect to

fuel consumption compared to an SI engine of similar size and build. So why is it that passenger

cars with diesel engines have been rather unpopular for many years?

One major drawback of the diesel engine is its low power-density with respect to size and weight

due to the lean operation. The amount of fuel that can be injected into the cylinder without creating

too much smoke is limited (typically to 1.3). However, the power-density can be increased by

super- or turbocharging the engine, i. e. forcing more air into the cylinder so that more fuel can be

burnt in the same volume without increasing the air to fuel ratio (AFR). This is typically achieved

with a variable geometry turbocharger (VGT) nowadays. Its turbine is placed in the exhaust stream

and utilises the energy of the exhaust gas to compress the air in the intake manifold via a compressor

on a connected shaft. A common type of VGT is the variable nozzle turbine (VNT) turbocharger

where the turbine is equipped with guide vanes which are adjustable by an actuator so that their

angle can be set according to the mass flow of the incoming exhaust gas. Compared to conventional

turbochargers with fixed geometry and waste-gate (turbine by-pass to dispose energy which would

otherwise cause damage at high speed) or detuning, the VGT allows a greater control authority to

achieve good boost at all engine operating set-points without sacrificing efficiency and dynamic

behaviour [Moo87].

Another disadvantage of the diesel engine is its typical combustion generated noise. It is due

to the initial rapid heat release immediately following the ignition-delay period where a significant

quantity of premixed charge exists. This can be avoided by injecting a small amount of fuel prior to

the main injection which smoothes the heat release. This so-called pilot injection has only become a

practical possibility with recent modern injection systems. Alternatively, injection rate modulation

can be applied with a modern common rail injection system to achieve lower combustion noise

[RGG00]. Due to this development, a modern diesel engine can hardly be distinguished from a

gasoline engine by its noise emissions once it has warmed up.

An increasingly important topic with regard to internal combustion engines is their emission

performance due to increasing environmental awareness and legislation as already mentioned. The

1.1. Background 5

regulated pollutants in the exhaust of internal combustion engines consist of unburnt hydrocarbons

(uHC), oxides of nitrogen (NO and NO2 , commonly referred to as NOx ), carbon monoxide (CO),

and particulate matter (PM; mainly soot, also absorbed hydrocarbon compounds). Reactions of HC

and NOx in the presence of sunlight cause photochemical smog (ground-level ozone) which initiated

the first emission standards for automobiles in California in the early 1960s. CO is toxic and the

reason why cars should not be left running in closed spaces such as garages. Large particulates

reduce visibility and some of the HC compounds are actually mutagenic. Recent research suggests

that the invisible small particulates (ultrafines) cause more damage to the human body than the

visible smoke (cf. Section 2.5). Furthermore, passenger cars are one of the major sources of the

green-house gas carbon dioxide (CO2 ) which is subject to legislature in some countries via fuel

economy standards or CO2 taxes.

The emission characteristics of CI and SI engines are quite distinct due to the different combustion

processes. The overall raw gaseous pollutant emissions of the diesel engine are actually considerably

lower than for the gasoline engine. However, due to its lean operation, the CI engine cannot be

operated with a three-way catalyst (which requires stoichiometric air-fuel mixtures) to efficiently

purify the exhaust gas. Therefore, NOx is present in larger quantities in the tailpipe of the diesel

engine, although the feedgas concentrations are comparable. The other main pollutant is particulate

matter which is specific for diesel engines (and lean-burn gasoline engines) where the inhomogeneous

air-fuel mixture can be locally very rich. It is not true to say that the conventional gasoline engines

emit no particulate matter, but the levels are very low indeed and legislation seems to be moving in

some countries to standards where diesel engines will be required to meet standard gasoline engine

levels. On the other hand, the high level of oxygen available during the CI combustion process

renders the CO emissions insignificant. Diesel HC emissions are about five times lower than in SI

engines but still significant (oxidation catalysts are typically used to oxidise unburnt HC emissions).

Finally, the lower fuel consumption of the diesel engine results in lower CO2 emissions.

Intake Manifold

EGR

Valve

Intercooler

Cooler

Compressor Turbine

VGT

Oxicat

6 Chapter 1. Introduction

The currently available techniques to reduce the emission of the two main pollutants, namely NOx

and particulate matter, are discussed in detail in Chapter 2. At this point, it shall only be mentioned

that the recirculation of exhaust gas into the cylinder (EGR) has been proven to be very efficient in

NOx reduction. The rather complicated mechanisms behind this effect will be investigated in Section

2.2. External EGR (recirculation of the exhaust gas via an EGR valve into the intake manifold) has

become standard equipment of production type cars and is also a feature of the four-cylinder common

rail diesel engine under consideration in this dissertation (cf. Figure 1.2). Refer to Appendix B for

a more detailed description of the engine facilities used in this project.

Figure 1.2 shows the configuration of a diesel engine including a turbocharger (VGT) and an

exhaust gas recirculation (EGR) system. In order to avoid heating up of the inlet charge, an EGR

cooler is used to cool the recirculated exhaust gas down. One interesting feature of the set-up in

Figure 1.2 is that both the EGR and VGT are driven by the exhaust gas and are therefore strongly

coupled [NKM00]. Hence, the airpath of the engine with these actuators constitutes an inherently

multivariable control problem.

The objective of control engineering is to manipulate a process such that it varies with time in a

prescribed manner. If there were no disturbances and every signal was known exactly, feedforward

control by inverting the plant properties would be ideal. However, noise and disturbances are always

present in physical systems. To handle this problem, the concept of feedback has been introduced.

A basic feedback structure is shown in Figure 1.3.

Desired

Output Error Control Output

Response Signal Output Measurement

Controller Plant

The controller computes a control output based on the error signal which is the difference of

the reference input and the measurements. This output is applied to the plant via actuators in order

to achieve the required behaviour. In this feedback framework, the controller now has information

about disturbances from the measurement and it can react to the error signal according to a specifiable

control law.

The signal paths in Figure 1.3 can also represent vector signals. This enables us to handle

multiple-input multiple-output (MIMO) systems in the same framework. Control design method-

ologies for MIMO systems were developed after the introduction of the state-space description

1.2. Outline 7

(which describes linear systems as a set of first order differential equations in a particular form)

at the beginning of the 1960s. The coordinated control of VGT and EGR in a diesel engine is

one example of a highly coupled multivariable control problem. It has been shown in [NMK+98]

that a multivariable controller could yield better performance than single-input single-output (SISO)

techniques in low speed/low load regions which results in a more rapid acceleration.

The design of a controller is usually based on a mathematical model of the plant. However,

mathematical models never represent nature perfectly. Although the feedback structure allows for

some modelling error, it is still a leap of faith from simulation to application of a control system

for the engineer. Robust control theory was introduced to explicitly address this problem. It gives

the engineer the opportunity to handle plants which are uncertain. For example, the mathematical

model of a diesel engine is very complex and hence has to be suitably simplified for control design

[GA98] which introduces uncertainty. The robust control design methodology then optimises the

controller for this set of plants. There is still no way to guarantee that the actual plant lies in the set,

but the leap of faith required from the user is certainly reduced.

The robust control theory considered in this report is based on the H norm of the transfer

matrix of linear systems. The H norm was first introduced by Zames [Zam81] in 1981. It is a

measure of the largest energy gain of a (multivariable) system and is therefore suitable for describing

worst-case scenarios in the presence of uncertainty. Computationally tractable state-space solutions

for H robust control problems became available at the end of the 1980s [DGK+89] and software

packages have been developed for the controller synthesis [BDG+94]. The application of H robust

control requires the problem to be set up in the so-called general framework which will be described

in Section A.1.6. The design methodology not only results in a controller but also gives the engineer

an indication of how much uncertainty can be tolerated.

The benefit of robust control for engine design is twofold. Firstly, diesel engines are highly non-

linear, hence their models are usually too complicated for standard control design. The conventional

way of handling this situation is to linearise the plant, which introduces approximation errors that are

usually neglected for the design process. Robust control gives the opportunity to explicitly account

for this error using an appropriate uncertainty description. Secondly, the engine control system is

optimised on a test facility engine and then applied to an entire range of production type cars which

inevitably inherit uncertainty due to their variability. Robust controllers are actually guaranteed to

yield the performance achieved on the uncertain model set defined by the engineer on any plant that

lies in this set.

While the conventional robust control tools require the linearisation of the engine model at dif-

ferent operating points, algorithms based on linear matrix inequalities (LMIs) have recently emerged

which can handle the robust control of linear parameter-varying (LPV) plants. However, the refor-

mulation or approximation of a nonlinear model as an LPV model is in general nontrivial.

1.2 Outline

After this brief introduction to internal combustion engines and control, the remainder of this dis-

sertation will be organised as follows:

8 Chapter 1. Introduction

This chapter presents a review of diesel engine emissions and current solutions to reduce the levels

of oxides of nitrogen (NOx ) and particulate matter (PM) in the tailpipe. The subsequent work

on modelling and control of the airpath of the diesel engine is ultimately aimed at improving these

emissions, since state-of-the-art aftertreatment (especially NOx traps) require precise control of EGR

and VGT.

This chapter gives an overview of mathematical mean-value modelling of the airpath of diesel engines

and provides the derivation of a nonlinear mean-value model for control design with emphasis on

the turbocharger parameterisation. Time-domain validation data are given for an initial assessment

of the model accuracy.

Firstly, the EGR-VGT control problem is stated and the effects of linearising the model at a fixed

operating point (1500 rpm, 85 Nm) are discussed. Secondly, the properties of the airpath model and

plant are investigated based on the step responses at the chosen operating point. Finally, model

validation in the frequency domain leads to the identification of the main uncertainties in the model.

A design study of robust H loopshaping controllers for the EGR-VGT control problem at the chosen

operating point is presented. A two degrees of freedom approach yields satisfactory performance.

Replacing the general coprime factor uncertainty description by more application tailored uncertainty

parameterisations leads to improved performance. The evaluation of the different controllers is based

on simulation as well as experimental results.

In order to allow for a control design on a wider operating range, a low-order linear parameter-

varying (LPV) model of the airpath is derived. A comparison to the nonlinear model from Chapter

3 indicates that it might be useful for control design despite its limited complexity.

Based on the LPV model, an H robust controller is synthesised using linear matrix inequalities

(LMIs) and evaluated in simulation and experiments. The results prove that this approach is suitable

for the EGR-VGT control over a wide range of the speed-load operating regime of the engine. As

opposed to other gain scheduling methods, the LPV controller comes with robustness guarantees.

1.2. Outline 9

Chapter 8: Conclusions

The conclusions and contributions of the presented work are discussed and potential areas for further

work are highlighted.

Appendices

The background of robust control required for this dissertation can be found in Appendix A, while a

description of the engine test facilities is provided in Appendix B. Parameter values for the engine

and the model are given in Appendices C and D.

10 Chapter 1. Introduction

2

Diesel Engine Emissions and Aftertreatment

The diesel engines reputation as a noisy, smoky, and sluggish power plant (the reasons for this have

been explained in Chapter 1) has changed due to modern diesel engine technology which allows one

to combine the inherent low fuel consumption with excellent driving performance and low emission

characteristics [LSG99]. After carbon dioxide (CO2 ) was identified as a greenhouse gas contributing

to global warming, diesel engines have emerged as an alternative to gasoline engines due to their

low fuel consumption and hence low CO2 emission. While carbon monoxide (CO) emissions are

negligible in CI engines due to lean operation and emissions of unburnt hydrocarbons (uHC) can be

handled with oxidation catalysts, the emissions of oxides of nitrogen (NOx ) and particulate matter

(PM) are of particular concern and therefore the topic of this chapter. With respect to aftertreatment

solutions, the focus will be on NOx traps because their regeneration requires tight control over the

airpath and constitutes the main motivation for this dissertation.

The formation of NOx as well as PM is closely linked to the combustion process (cf. Section

2.1) which depends on engine design variables such as combustion chamber and fuel injector design,

pressure and timing of the injection (modern injection system such as common rail also allow multiple

injections), swirl ratio, valve timing, compression ratio, etc. In general, these variables can only be

optimised for the reduction of one of these two main pollutants due to the so-called NOx -PM trade-off

described in Section 2.1 (although there are exceptions, e. g. two-stage combustion can break the

trade-off by forming a fuel rich mixture at the initial combustion stage to prevent NOx formation

and inducing strong turbulence in the combustion chamber at the later stage of the combustion

to oxidise the particulates [CA96]). Typically, the other pollutant is controlled by aftertreatment.

Alternatively, NOx and PM aftertreatment solutions can be integrated which allows the combustion

system and engine calibration to be optimised with respect to performance/efficiency. This chapter

gives an overview of currently available aftertreatment techniques. Concerning in-cylinder emission

reduction techniques, only exhaust gas recirculation (EGR) and fuel composition (especially water

emulsions) will be discussed. Engine design changes are a whole other area to address emissions

and performance, but they are beyond the scope of this review.

Legislation is the driving force for the reduction of diesel engine emissions as already mentioned

in Chapter 1. Figure 2.1 depicts the current and future emission standards for light-duty diesel engines

in the European Union. The depicted NOx -PM trade-off curve for a 1590 kg vehicle with a 2.0 litre

high-speed DI diesel engine indicates that advanced aftertreatment is not necessary to satisfy Euro

IV regulations (at least in this weight class) [Joh00]. However, marketing reasons require further

11

12 Chapter 2. Diesel Engine Emissions and Aftertreatment

reduction of both NOx and PM in Europe, while the proposed emission standards in the US are

already tighter. Since a direct comparison between European and US standards is not fair due to

completely different emission drive cycles, Figure 2.1 only shows the European regulations.

0.15

Euro I (NOx +HC) Euro I from 1992

Euro II from 1999 for DI

Euro III from 2000

Euro II (NO x +HC) Euro IV from 2005

0.10

[PM] = g/km

Euro III

0.05

0.5 1.0

[NOx ] = g/km

Figure 2.1: Evolution of emission standards in the European Union and typical NOx -PM trade-off

curve for a light-duty vehicle [Joh00].

After describing the formation of NOx and PM during the diesel combustion process in the next

section, this chapter will give an overview of diesel emission reduction strategies. Exhaust gas

recirculation (EGR) is not only a well established means to reduce NOx emissions, it also plays a key

role in the regeneration strategy for NOx traps. Therefore, its different effects on NOx emissions are

discussed in Section 2.2. The emphasis of this chapter is on NOx traps in Section 2.3. In addition

to their operating principle, the status quo of regeneration and desulphation techniques which are a

crucial factor in the potential success of NOx traps will be presented. Section 2.4 gives an overview

of alternative solutions for NOx reduction such as DeNOx catalysts, selective catalytic reduction

(SCR), and non-thermal plasma technology, while Section 2.5 elaborates on the nature of diesel

particulates and their elimination in the diesel exhaust by particulate filters. The chapter finishes

with a discussion of integrated NOx and PM solutions in Section 2.6 and conclusions in Section 2.7.

It is worth warning the reader that this topic is still a very active research area and it is not clear

yet which aftertreatment solutions will prevail. Therefore, this review describes the most promising

methodologies up to date.

The diesel engine combustion begins around 5 after start of injection (ASI) in a locally rich premixed

zone as described in Section 1.1.1 and depicted in Figure 1.1. The products from this early reaction are

small, partially burnt fragments of hydrocarbons which are generally believed to lead to the formation

of polycyclic aromatic hydrocarbons (PAH) that constitute the building blocks for particulates in

2.1. Formation of Nitrogen Oxides (NOx ) and Particulate Matter (PM) 13

flames [FDH+99]. When the air entrained during the premixed burn is entirely consumed, the

temperature reached by this partial oxidation is about 1600 K and the reaction products (CO, CO2 ,

fuel fragments and water vapour) will subsequently receive heat from the hot diffusion flame to form

an environment favourable for the formation and agglomeration of soot particles. The fuel fragments

and soot particles are subsequently transported through the interior of the plume toward the boundary

of the diffusion flame sheath where they are oxidised in a thin reaction layer at temperatures around

2,700 K. Experiments show that these particulates are completely consumed by the hot diffusion

flame sheath so that diesel soot emissions appear to be the result of quenching this final phase of

oxidation. The temperature profile of the quasi-steady flame is shown in Figure 2.2.

The major part of NO from the diesel combustion is formed by the oxidation of atmospheric

nitrogen via the extended Zeldovich mechanism [Hey88]:

O + N2 NO + N (2.1)

N + O2 NO + O (2.2)

N + OH NO + H (2.3)

This formation of NO is highly temperature dependent due to large activation energies for the forward

reaction (2.1) and the reverse reactions in (2.2) and (2.3). It is initiated at temperatures greater

than 1900 K and the NO stays around during cooling since the reverse reaction is very slow. Fuel

nitrogen is also a source of NO via a different and yet to be fully explained mechanism, however,

its contribution is less significant in diesel engine combustion than the formation of NO via the

Zeldovich mechanism. The produced NO can then be converted to NO2 by reactions such as:

NO2 + O NO + O2 (2.5)

In the preceding discussion of the diesel combustion process, high temperatures in a region where

oxygen is available along with nitrogen characterise the diffusion flame sheath rendering this area

14 Chapter 2. Diesel Engine Emissions and Aftertreatment

the ideal region for the formation of NOx . According to [FDH+99], two-thirds of the total NOx

emissions are formed in the diffusion flame while one-third is formed in the hot post-combustion

gas regions. Chemical equilibrium considerations indicate that at typical flame temperatures, the

NO2 /NO ratio should be negligibly small which is true for SI engines. However, in CI engines, NO2

can contribute 10 to 30 % of the total NOx emissions because the back reaction from NO2 formed in

the flame to NO (2.5) is quenched by mixing with cooler regions especially at light load [Hey88].

Soot particles, which heavily contribute to the total mass of particulate matter emitted by the engine,

are formed in the cylinder in the locally rich regions of the inhomogeneous combustion. The subse-

quent soot burn-up at the boundary of the diffusion flame sheath is favoured by high temperatures.

However, high temperatures also favour the generation of NOx . Hence, decreasing the combustion

temperature (e. g. by exhaust gas recirculation as explained in Section 2.2) for lower NOx emissions

results in an increase of PM. This dilemma is known as the NOx -PM trade-off.

The recirculation of exhaust gas through an EGR valve into the intake manifold where it dilutes the

incoming fresh air (thereby replacing oxygen with carbon dioxide and water vapour in the charge) is

a well established and efficient means of reducing in-cylinder NOx emissions. EGR is also a feature

of the engine under investigation as depicted in Figure 1.2. However, the reduction of NOx occurs

at the expense of (initially slightly) lower thermal efficiency and higher PM emissions (following

the NOx -PM trade-off). In order to understand the principles behind the combustion process with

a diluted charge, Ladommatos et al. separated five possible effects of EGR (here presented in the

order of significance for CI engines) on the in-cylinder NOx production:

1.) Dilution Effect. The replacement of in-cylinder oxygen with exhaust gas reduces the availability

of oxygen during the combustion, which lowers the temperature of the whole combustion process

in the premixed burn as well as in the diffusion flame, where most of the NOx is created according

to the combustion model from Section 2.1. This is due to the effect that the combustion still occurs

in a region where the ratio of fuel and oxygen is around stoichiometric but now with exhaust gas

acting as a dilutent. The NOx emissions decrease significantly with reduced oxygen availability

and its consequent reduction of the peak flame temperature. The latter has actually the dominant

effect on the NO formation through the extended Zeldovich mechanism in (2.1) - (2.3). Note that the

dilution effect is the most influential one in the NOx reduction by exhaust gas recirculation. While the

dilution of the inlet charge has only a minor effect on the thermal efficiency of the engine, particulate

emissions increase significantly with decreasing inlet charge oxygen. ([LAZ+96],[LAZ+97b])

2.) Inlet Charge Temperature Effect. The recirculation of exhaust gas raises the temperature of the

inlet charge and consequently leads to a higher flame temperature resulting in higher NOx emissions

2.3. NOx Traps (Adsorbers) 15

(although less significantly than the dilution effect). It is observed that an increase in inlet charge

temperature results in substantially higher emissions of PM (especially soot). This could be due to

an increased rate of fuel pyrolysis in the premixed burn. The thermal efficiency of the combustion

decreases slightly with rising inlet charge temperature which can be attributed to higher heat losses.

In order to minimise the inlet charge temperature effect, the recycled exhaust gas needs to be cooled.

([LAZ+96],[LAZ+97b])

3.) Ignition Delay Effect. While the increase of the inlet charge temperature shortens the ignition

delay, the lower oxygen availability increases it such that the net effect of EGR is an increase of the

ignition delay. This increase potentially affects the combustion by increasing the fuel fraction burnt

during the premixed burn and by shifting the auto-ignition towards the combustion chamber wall,

however, the lower oxygen availability due to the dilution effect actually reduces the peak rate of

premixed burning. This reduced burning rate and the delay of the combustion towards the expansion

stroke reduces peak temperatures and pressures and the time the combustion gases spend at these

conditions. Thereby, NOx emissions are reduced (although significantly less than the dilution effect)

at the expense of PM emissions and thermal efficiency. Obviously, the ignition delay effect can be

compensated for by adjusting the injection timing. ([LAZ+96],[LAZ+98])

4.) Chemical Effect. This effect describes the phenomenon that the dissociation of burnt gases

at high temperatures can result in the reaction of NO with the resultant hydrocarbon free radicals

(so-called denoxing). However, it is suggested that other mechanism might also play a role. While

approximately 10 % of the total NOx reduction can be attributed to the chemical effect of carbon

dioxide, the chemical effect of water vapour was negligible (the dissociation of water is highly

endothermic so that it occurs only at higher temperatures than the dissociation of carbon dioxide).

([LAZ+96b],[LAZ+97],[LAZ+97b])

5.) Thermal Effect. Due to the higher average specific heat capacity of the reacted exhaust gas

components as compared to the specific heat capacity of oxygen that is replaced in the charge the

combustion temperature is lowered. Traditionally, this was thought to be the reason for the NOx

reducing effect of EGR. However, in diesel engines where the exhaust gases replace atmospheric

air, the change in the heat capacity of the charge is minor. Thus, the thermal effect is considered

insignificant here. Note that this is different from conventional SI engines, where the recycled exhaust

gases are added to a virtually unchanged fuel/air mass (i. e. the manifold pressure is higher with an

EGR flow). Hence, the thermal capacity of the cylinder charge is increased due to the increase in

both mass and specific heat capacity which renders the thermal effect the most significant one in SI

engines. ([LAZ+96b],[LAZ+97],[LAZ+97b])

The raw emissions levels of NOx for lean-burn engines and stoichiometrically operated engines are

actually quite similar. However, in SI engines running stoichiometric, NOx can effectively react

16 Chapter 2. Diesel Engine Emissions and Aftertreatment

with uHC and CO to nitrogen gas using a three way catalyst while preferring to react with excess

oxygen in lean environments. Different solutions have been developed over the years [Joh00, Joh01]:

NOx Traps (NOx Adsorbers) which are described in this section and DeNOx Catalysts, Non-thermal

plasma treatment as well as Selective Catalytic Reduction which are described in Section 2.4. While

these are all aftertreatment solutions, exhaust gas recirculation is an important way of reducing NOx

emissions by changing the combustion characteristics as described in detail in Section 2.2.

The NOx trap technology had first been introduced for direct injection gasoline engines and was

then adapted to diesel applications which run at lower exhaust gas temperatures due to the high

air mass flows in lean operation and high thermal efficiency of the diesel combustion. Hence the

temperature window of NOx traps for diesel applications has to be lowered as depicted in Figure

2.4. Diesel NOx traps can already achieve over 80 % efficiency at steady-state, but have problems

with sulphur poisoning (desulphation of the trap requires high temperatures which expedite thermal

aging), and also with the fuel penalty due to regeneration in a rich environment.

The operating principle, which is depicted in Figure 2.3, was discovered while investigating the

selective reduction of NOx by uHC (cf. Section 2.4.1) and is described in Toyotas milestone paper

in 1995 [MMK+95]: The trap temporarily stores NOx as a nitrate on alkaline earth compounds under

lean conditions. In periodic rich conditions, called regeneration, which last for a few seconds after of

the order of 100 seconds of lean operation, the nitrate dissociates and the NOx is reduced as in a three

way catalyst. The reduction occurs in a rich environment as opposed to DeNOx catalysts, hence,

the mechanism is advantageous for less generation of nitrous oxide (N2 O, a potent greenhouse gas,

cf. Section 2.4.1). Note that the regeneration, i. e. running a diesel engine rich, is a major challenge

and is only possible with state-of-the-art diesel injection systems such as common rail to avoid

excessive black smoke.

H2 O,CO2 , HC,CO,H2

NO+O2 N2

NO3 NO3

NO2

Pt Ba Pt Ba

Al2 O3 Al2 O3

In NOx storage catalysts, both oxidation steps occur in the same device (NO to NO2 on a noble

metal substrate, e. g. platinum, and NO2 to nitrates in a second step). The trap only works in a

2.3. NOx Traps (Adsorbers) 17

temperature window (typically 250 to 500 C, cf. Figure 2.4); the lower limit is determined by the

oxidation of NO to NO2 (slow at low temperatures at the noble metal surface), the upper limit is

determined of the thermal stability of the nitrates (they become unstable). Due to the use of high

EGR rates at low loads, the low activity at 100 to 250 C comes along with relatively low NOx

flow emissions. The increase of activity at higher temperatures is accompanied by higher NOx

concentration and space velocity. In [GBB+00], different adsorber materials exhibit quite different

temperature windows and one trap achieved 85 % NOx efficiency on the NMVEG (New Motor

Vehicle Exhaust Gas) cycle while three other traps only achieved around 45 % (however, they all

display similar behaviour towards regeneration efficiency; especially interesting is the lack in the

low temperature region regenerability, important under typical urban driving conditions, which is

ascribed to a slow NOx reduction at that temperature compared to the nitrate decomposition rate).

100

90

Diesel Application

[NOx Conversion] = %

80

Gasoline Application

70

60

50

40

30

20

10

0

150 200 250 300 350 400 450 500

[Temperature] = C

Figure 2.4: Typical NOx trap temperature window for a diesel (solid) and lean gasoline (dashed)

application obtained from model gas tests [KBH+01].

In addition to the temperature, other operating parameters influencing NOx adsorption efficiency

are the trap regeneration frequency, the space velocity, and the NOx feed-gas concentration. With

respect to the trap formulation, it has been shown in [MMK+95] that increased storage component

basicity increases the NOx storage potential. However, it deteriorates the activity of the three-way

catalytic reduction. Hence, a trade-off has to be found. In the same paper, it is observed that greater

particle size of the noble metal reduces its activity. In a NOx trap this behaviour is particularly

distinctive for NOx conversion as the storage amount of the adsorber increases when noble metal is

in the proximity which is less likely if the metal particle size increases.

In [BBM+96] the effect of a closed-coupled oxidation catalyst (for faster light-off in lean gasoline

applications) upstream of a NOx trap is investigated. The minimum regeneration duration had to be

increased from 300 ms to 700 ms due to the oxygen storage and hence oxidation of the reductants

for the NOx trap. It is observed that the light-off temperature for lean NOx conversion is reduced.

This is explained by the oxidation of uHC and CO in the oxidation catalyst under lean conditions

18 Chapter 2. Diesel Engine Emissions and Aftertreatment

which reduces their partial pressure and therefore lowers the competition for the platinum sites

on the trap leading to improved NO oxidation. The authors do not mention the NO oxidation in

the oxidation catalyst though. Moreover, it is observed that the close-coupled catalyst improves

the sulphur tolerance of the system significantly which is attributed to a largely reversible sulphur

storage mechanism on the close-coupled catalyst.

While the formulation of NOx traps has already been thoroughly investigated, the work is now

shifting towards reducing the impact of NOx trap regeneration/desulphation on vehicle drivability

especially during transients.

In order to avoid the breakthrough of CO and uHC emissions and a higher than necessary fuel penalty,

the dosage of the rich spikes has to comply with the amount of NOx adsorbed in the trap. Hence, a

precise observation of the trap state has to be implemented. Moreover, the rich excursion needs to

be imperceptible to the driver.

An enriched exhaust gas composition can be achieved either via engine modifications: 1.) Re-

duction of fresh air (intake air throttle, high EGR rate), which is rather slow, 2.) Increase of fuel

mass (very early or late injection timing, changing the injection characteristic), 3.) Post injection or

via exhaust system modifications: Injection of reducing agents in front of the trap.

Due to the inhomogeneous mixture formation in a diesel engine, rich operation leads to locally

enriched zones, which usually results in a significant increase in black smoke emissions. However,

an optimised calibration of all engine operating parameters can reduce the increase of black smoke

to an acceptable level. In [LSG99], uHC concentrations up to 7,500 ppm and CO concentrations as

high as 4 % are reported to be achieved in steady-state, which is sufficiently high for NOx adsorber

regeneration, without exceeding a smoke number of 2.8 Bosch. Moreover, a test cycle is shown to

be successful where every 50 s of lean operation, a rich spike for 2 s would occur. Typically, during

the rich spikes, sharp NOx desorption peaks occur. This happens when the nitrates dissociate in a

rich environment but not enough reduction capability is available to reduce all of the released NOx .

As described in Section 2.2, decreasing the AFR by increasing the EGR rate results in higher

smoke emissions. However, Akihama et al. recently discovered that this holds true only up to a

critical point [ATI+01]. Further increasing the EGR level then leads to a sharp decrease of smoke

emissions even in rich conditions. In fact, total smoke is high, but it is largely an unburnt fuel

cloud, rather than black (carbonaceous) soot. This phenomenon can be explained by the -T-map,

where =1 is the local fuel to air equivalence ratio. Figure 2.5 shows the local temperature and

equivalence ratio region where soot is generated indicating that soot formation is suppressed at

temperatures below 1,700 K even if rich combustion occurs. Hence, smokeless rich combustion is

achieved by reducing the temperature by very high EGR rates such that the soot formation region

is avoided. No improvement of the mixture formation is required. The lower temperature freezes

the reactions from PAH (polycyclic aromatic hydrocarbons, cf. Section 2.1) to soot particles. Thus,

while the soot emissions are reduced, the PAH emissions which account for the soluble organic

fraction (SOF, cf. Section 2.5) of the PM emissions increase. However, the latter can be oxidised by

2.3. NOx Traps (Adsorbers) 19

a catalyst. Due to the large amount of EGR (about 55 %) the rich operation is limited to low load. It

has the advantage of increasing the exhaust gas temperature under these conditions to values similar

to gasoline engines, thus allowing improved catalytic removal of uHC and CO.

10

9

Equivalence Ratio =1

7

6

5

Smokeless

4

Rich

3 Combustion Soot

Generating

2 Region

1

Increasing EGR

0

1000 1400 1800 2200 2600 3000

[Temperature] = K

The sulphur content of diesel fuel is typically higher than in gasoline, and the sulphur content of the

respective mineral oil depends greatly on the location of the oil fields. Typical diesel fuel in Europe

has a sulphur content of 370 ppm (the legal limit is 500 ppm since 1996), but 10 ppm diesel fuel is

nowadays also available in many markets. Recently, the Environmental Protection Agency (EPA) in

the US proposed a 15 ppm limit on sulphur content starting in June 2006. However, current research

indicates that even with such low sulphur contents desulphation strategies for NOx traps are required.

The effects of sulphur in diesel fuel can be summarised as follows [KAD+99]:

1. Emissions: Formation of sulphur dioxide SO2 during combustion with by-products SO3

and H2 SO4 (involved in causing acid rain); formation of sulphates and adsorption on soot-

particulates.

2. Corrosion and wear: Increasing wear of engine parts through the corrosion by internally

formed acid under cold start conditions; corrosion in the exhaust system through sulphuric

acid and condensate accumulation.

3. Formation of sulphates due to oxidation of sulphur in the catalyst and therefore increasing PM

(in particular at high exhaust gas temperatures, which supports the finding in [SMB00], where

PM number increased at high speed).

20 Chapter 2. Diesel Engine Emissions and Aftertreatment

In the NOx trap, the oxidation of SO2 to SO3 is in direct competition to the oxidation of NO to

NO2 . The formation and accumulation of sulphates in the storage catalytic converter also occurs in

a similar way to the nitrate process. Hence, the trap becomes poisoned by sulphates. Unfortunately,

the sulphates are thermodynamically more stable than the nitrates, which means that the reaction

conditions for sulphate regeneration are much more extreme (even under rich exhaust gas conditions,

temperatures higher than 650 C and for a longer period are required).

In 1999, DaimlerChrysler investigated the effect of the fuel sulphur level on the performance

of a NOx storage catalyst [KAD+99]. Their experiments show that using standard diesel fuel with

a sulphur content of 370 ppm, the efficiency of the NOx trap is significantly reduced within a few

operating hours. However, with 10 ppm diesel fuel, over 50 % NOx reduction efficiency can be

achieved on the New European Drive Cycle (NEDC). Note that for a sulphur fuel concentration of

less than 50 ppm, the oil contribution has to be taken into account such that desulphation strategies

are required for NOx traps even with <5 ppm sulphur content fuel [GBB+00]. In this paper, Guyon

et al. also show in their experiments that the NOx conversion efficiency drops from 95 to 55 % after

4,250 km without desulphation using 50 ppm fuel. Moreover, desulphation (=0.83, T=680 C for 10

minutes) does not allow one to recover the initial efficiency: After five desulphations corresponding

to 12,000 km, 30 % of the initial NOx storage capacity had irreversibly been lost. It is worth

mentioning that results of this type are dependent on details of the NOx trap formulation constant

improvements are being made, but the basic challenges as evidenced by the results being quoted

seem universal.

A particularly thorough investigation of desulphation strategies has been conducted by Klein et

al. from dmc2 and AVL [KBH+01]. The authors show that desulphation in a vehicle is practical. In

order to demonstrate this, a diesel NOx storage catalyst has been developed which has a significantly

higher NOx conversion rate between 150 and 250 C as compared to NOx traps for lean-burn gasoline

engines which have typically higher exhaust gas temperatures (cf. Figure 2.4). The conversion rate

at different temperatures is investigated using a model gas at different temperatures but fixed space

velocity (which does not represent typical engine installations) and averaging over 28 lean-rich

cycles. For sulphation of this trap, a diesel engine is run at a fixed operating point for 120 s lean and

5 s rich (=0.90) using 330 ppm sulphate content in the fuel. A fresh catalyst reaches an average NOx

conversion of 85 % which reduces to 63 % after ten desulphations. However, one has to consider

that the lean-rich cycle time has been optimised for a fresh catalyst and could be adapted to the state

of the catalyst for better performance. This would hence require one to observe the catalyst state

in the vehicle. The desulphation temperature is also examined: It turns out that for a practicable

desulphation (low sulphate content in the trap after desulphation) temperatures higher than 600 C

are necessary which, however, permanently reduces the NOx conversion rate at low temperatures.

In tests with sulphur-free fuel, it is shown that the main reason for the reduced NOx conversion is

the thermal aging of the trap due to the high temperature desulphations and not sulphate poisoning.

Further investigations show that for an efficient desulphation, rich spikes of <0.9 and length >10 s

enable a sufficient desulphation within a few rich-lean-cycles.

As explained before, the desulphation process requires temperatures in the exhaust of 500 to

700 C (as compared to gasoline applications with trap formulations that allow desulphation at 500 C,

2.4. Alternative DeNOx Solutions 21

temperatures around 450 C required for DPF regeneration, and a temperature of 250 C for continuous

NOx reduction in the NOx trap). In [KBH+01], a temperature rise from 150 to 250 C (at 1300 rpm,

1 bar BMEP steady-state) is achieved using engine parameter variations (multiple injection, injected

fuel mass, injection timing, rail pressure, intake air throttling, and EGR rate), the rise up to 650 C is

then achieved using post injection which generates uHC for an exothermal reaction in the oxidation

catalyst (11,000 ppm uHC in feed-gas, double fuel consumption, soot and NOx emission less than in

standard calibration, nanoparticles not mentioned). The alternative of externally heating the exhaust

gas has not been examined.

The main problem with the described desulphation strategies is the permanent deterioration of

the NOx efficiency due to thermal aging. Moreover, spikes of bad smelling sulphur emissions (H2 S)

can occur during desulphation. An alternative solution is to place a sulphur trap in front of the NOx

trap as investigated by Strehlau et al. in [SLL+96] for a lean-burn gasoline engine. The SOx trap

adsorbs the sulphur oxides under lean conditions in order to keep them away from the NOx trap.

During the rich spikes for regeneration of the NOx trap, the sulphur trap releases the sulphate mainly

as SO2 which is then not oxidised to SO3 in the NOx trap (due to the rich environment) and hence not

stored as sulphates. This is shown to work in general, but it will mostly delay the sulphur poisoning,

thus extending the NOx adsorber lifetime, rather than preventing the poisoning completely.

2.4.1 DeNOx Catalysts

DeNOx catalysts consist of precious metal or zeolites and they use unburnt hydrocarbons (as opposed

to ammonia which leads to the SCR reaction, cf. Section 2.4.2) to reduce NOx in a lean environment.

For example, copper ion exchanged ZSM-5 is known to decompose NO into N2 and O2 . The exhaust

uHC is increased by in-cylinder injection modifications or direct injection of uHC into the exhaust

gas. DeNOx catalysts achieve in general up to 35 % efficiency, but have up to 6 % fuel penalty due to

the need of hydrocarbon reductant. Higher efficiencies (up to an impressive 70 %) have been shown

using diesel fuel reductant on engine exhaust, resulting in a fuel penalty of 4 % but with a catalyst

volume of four times the engine swept volume. The major problem is the tight temperature window

in which the catalyst works efficiently. For gasoline applications typical zeolite catalysts (high

efficiency at higher temperatures than desirable in diesel engines) exhibit, moreover, poor thermal

stability due to the H2 O present in the exhaust gas. Yet another problem is that NOx conversion in

an excess oxygen atmosphere using a catalyst of a noble metal family produces significant quantities

of nitrous oxide (N2 O), a potent greenhouse gas (which can be avoided by using NOx traps where

the conversion occurs near stoichiometric [MMK+95]).

In an SCR system, the NOx present in the exhaust gas is mixed with ammonia as a reductant and

catalysed to nitrogen and water. The ammonia reductant needed for the selective catalytic reduction

is typically obtained from on-board urea. The urea is injected into the exhaust gas. A typical SCR

22 Chapter 2. Diesel Engine Emissions and Aftertreatment

system consists of three different catalysts in series after the urea injection point [GSS+00]: 1.)

Hydrolysis catalyst which converts the urea with water selectively to ammonia (NH3 ) and carbon

dioxide, 2.) SCR catalyst (e. g. V2 O5 /WO3 /TiO2 ), where the ammonia reacts with NOx to form

nitrogen and water, and 3.) Oxidation catalyst to avoid ammonia slip during transient operation.

Over 90 % efficiency is possible with SCR systems. Problems are a possible ammonia slip,

the compactness of the system, and the requirement for an urea distribution network; an advantage

compared to traps is the less complex engine management requirement (however, a dosage system

for the urea is required). Recently, the advantages of first oxidising NO using an oxidation cat to NO2

have been discovered [GSS+00] for a V2 O5 /WO3 /TiO2 SCR catalyst. The higher NO2 /NOx ratio

results in better low-temperature efficiencies of the SCR (typically 30 to 40 %). However, at NO2

fractions larger than 50 %, the higher stoichiometry needed to convert NO2 (1.33 compared to 1.0

for NO), and the slower reaction of NO2 with NH3 which results in NH3 slip limit the maximum con-

version without secondary emissions. It has also been found in the same study that V2 O5 (vanadium

oxide) containing catalysts are irreversibly deactivated at temperatures above 700 C. In a follow-up

study, the same authors demonstrate the NOx reduction potential of ion exchanged synthetic zeolites

and their favourable aging properties (only slight NOx conversion reduction after aging at 800 C)

[GPS+01]. It is also shown that the crucial component to achieve high NOx conversion is NO2 ,

hence, a pre-oxidation catalyst is required.

There is some concern that high-efficiency SCR systems generate sulphate PM, therefore ultra-

low sulphur fuel is needed (although the SCR is not as sensitive to sulphates as a NOx trap). SCR is

effective in removing uHC and some of the PM (probably SOF, cf. Section 2.5). Since an SCR system

does not directly interact with the engine control unit, the technique is suitable for retrofitting. Urea

injection strategies are being reported. The first SCR system in a vehicle application was reported in

1995. The current research focuses on the development of solid reductant technology which would

allow to extend the refilling interval to service intervals.

While the SCR technology has been proven successful in stationary applications like engine

driven power stations for decades, the first studies on automotive applications emerged around 1995.

Havenith et al. [HVH+95] investigate an SCR system with urea injection and a downstream oxidation

catalyst in steady-state and transient operation. The steady-state efficiency decreases from above

90 % at low load to about 65 % at high load due to reagent oxidation. Note that below 25 % load,

urea injection is withheld due to low oxidation catalyst temperatures (300 C) where ammonia slip

might occur. Note also that the sulphate emission increases due to the oxidation cat and that the

N2 O emission is ten times higher than normally expected without NOx aftertreatment (ammonia

undergoes partial oxidation to N2 O and possibly more complete oxidation to NO and NO2 over an

oxidation catalyst). In transient cycles, the system is found to be effective in NOx reduction although

peaks in NOx emission occur after load steps, which is supposed to be caused by buffering effects.

NOx adsorbers with desulphation might draw 1 to 2 % more effective fuel penalty (considering

urea), but the user will have to deal with an added reagent on board. It is therefore likely that Europe

2.5. Particulate Matter 23

will commercialise SCR on a large scale and the US NOx traps. Certainly, NOx sensors will become

important and are already appearing on lean-burn cars in Europe.

The operating principle of the non-thermal plasma technology is as follows: A high local electrical

field is applied to generate microdischarges in the exhaust gas which directly produce highly energetic

electrons. Due to the short duration of the microdischarges (<100 ns) the electrons and the bulk gas

do not thermally equilibrate and, hence, the temperature of the bulk exhaust gas remains essentially

constant [LSG99]. The plasma generated electrons dissociate oxygen which reacts with unburnt

hydrocarbons to form a variety of partial oxidation products. Moreover, NO is oxidised to NO2 . When

uHC is present in the plasma the oxygen radicals react preferentially with the unburnt hydrocarbons

instead of with SO2 thus there is only a small conversion to sulphates as opposed to an oxidation

catalyst [Hoa01].

The gas phase plasma alone does not remove NOx from the exhaust, hence, a combination of

plasma and catalyst is needed. In [Hoa01] it is observed that a NaY catalyst achieves 58 % NOx

conversion at 180 C, a Cu-ZSM5 simply converted the NO2 back to NO. Therefore, the plasma-

catalyst NOx reduction obeys a different reaction mechanism at lower temperature than the thermal

catalysis.

The advantage of the non-thermal plasma technology is a wider temperature range for NOx

conversion (150 to 500 C) than non-plasma catalysts and reduced sulphur sensitivity (urea SCR and

zeolitic DeNOx do not extend to such low temperature, precious metal DeNOx does not extend to such

high temperature and NOx traps are more fuel sensitive). Moreover, in [Hoa01], high particulate

removal has been demonstrated although not yet simultaneously with NOx reduction in a single

system.

Non-thermal plasma technology achieves 60 to 70 % NOx reduction efficiency, but the fuel

penalty is around 6 % due to the cost of electric power and addition of unburnt hydrocarbons for

best conversion efficiency. However, note that these techniques are still in their early development

stages.

2.5.1 Nature of Diesel Particulates

Most of the particle number emitted by engines is in the nanoparticle range, D < 50 nm, while most

of the mass is in the accumulation mode, 50 nm < D < 1000 nm [Kit98]. Nanoparticles: Typical type

is uHC/Sulphate, but may also contain solid carbon and metal compounds; they form by nucleation

during dilution and cooling; they also appear in SI engines; there may actually be more in some state-

of-the-art diesel engines (the better combustion produces less accumulation mode particles on which

small particles can adsorb; this was observed on an engine with high values of the soluble organic

fraction SOF). Accumulation Mode Particles: Mainly carbonaceous soot agglomerates directly from

the combustion; 10 to 100 times more than in SI engines.

24 Chapter 2. Diesel Engine Emissions and Aftertreatment

Typical particle composition for a HDD engine on a transient cycle (in percent by mass):

41 % Carbon (Soot)

32 % Soluble Organic Fraction (SOF), consisting of 25 % unburnt oil and 7 % unburnt fuel

with high boiling point; the low boiling point unburnt components become uHC emissions

SOF values can range between 10 % and 90 % and are highest at light engine loads when exhaust

temperatures are low. The sulphuric acid/sulphate concentration is roughly proportional to the fuel

sulphur content. As the exhaust gas is cooled and diluted, nucleation, condensation, and adsorption

transform volatile materials to solid and liquid particulate matter.

The particle size distribution (Figure 2.6) shows three peaks corresponding to the following modes:

particle number

Accumulation Mode: 100 to 1000 nm diameter range, contains most of the mass

mode particles that have been deposited on cylinder and exhaust system surfaces and later

reentrained.

Current emission standards are the result of many studies that have shown a link between fine

particle (with a diameter of less than 2.5 micrometers) air pollution and adverse health effects

(e. g. [PTN+95]). These standards are based on mass and do not address particle size. However,

many recent studies indicate that at similar mass concentrations nanometer size particles are more

dangerous than micron size particles (e. g. [DLM98]). Due to their extremely small size, nanopar-

ticles (i. e. with a diameter of less than 50 nanometers) can penetrate deeper into the lung and are

more difficult to remove from the lung tissue. This is associated with respiratory inflammation and

acute pulmonary toxicity for people with a preexisting condition. Moreover, the large surface area

provided by nanoparticles in contact with the lung provides the opportunity for surface chemistry of

the particles to have a profound effect. The problem for legislation is that the mass is conserved, but

not the number. This phenomenon is due to nucleation and coagulation during dilution and sampling,

making it difficult to design a standard. Note that particulates also affect engine performance and

wear, e. g. when they are recycled to the engine via EGR.

The particle size distribution of solid soot particles generated in the cylinder varies little with

fuel. Soot mass is generated about 10 to 20 ATDC, with sizes up to 110 nm; as the piston

drops, these particles oxidise and decrease in number about two orders of magnitude and in

size to about 70 nm.

2.5. Particulate Matter 25

Number Weighting

Nanoparticles

D<50 nm

Normalised Concentration Mass Weighting

Ultrafine Particles

D<100 nm Fine Particles

D<2.5 m

PM10

D<10 m

Nuclei

Mode Accumulation Coarse

Mode Mode

[Diameter] = m

Nucleation and growth phenomena are strongly dependent on concentration (exhaust dilution

ratio), time, and temperature and, even soot concentration.

Diesel Particulate Filters (DPF) have been considered as a solution to the PM problem for twenty

years. Although not really necessary in Europe to satisfy Euro IV, much effort is being done on

developing filters in Europe, lead by Peugeots introduction of a DPF in serial production in 2001.

A typical filter is the ceramic wall-flow filter, where the exhaust gas is forced to flow through

the walls of the monolith which removes PM very efficiently. To avoid clogging up of the filter, the

PM has to be oxidised, either periodically or continuously. Direct oxidation by O2 is accomplished

at temperatures of around 550 C, where the carbonaceous soot oxidises. However, the exhaust

temperature of the diesel engines under normal operation is in the range of 150 C to 200 C, therefore,

efforts are being made to reduce the balance point temperature (BPT). The BPT is the temperature at

which particulate accumulation is equal to particulate oxidation, wherein back-pressure is constant

and the system is in balance. The BPT depends on flow rate, particulate composition, NOx content,

sulphur levels, soot loading rate, and perhaps many other engine/fuel parameters. The idea is to

bring the BPT down to ease regeneration see next section.

Filters are most effective at removing solid accumulation mode particles (letting through mainly

SOF and sulphates which are almost completely gaseous at typical filter temperatures), while cat-

26 Chapter 2. Diesel Engine Emissions and Aftertreatment

alysts remove mainly organic compounds that will contribute to SOF and nuclei mode [Kit98].

According to [Joh01], however, numerous studies have demonstrated that filters take out more than

90 % nanoparticles, but it might be that gaseous nanoparticle precursors are not removed and sulphate

nanoparticles might actually increase.

There are actually conditions, where filters increase the number of nanoparticles, since they take

away the soot under hot conditions, while letting escape gaseous aerosol precursors which condense

downstream in larger numbers than if carbon soot was available.

Three things have to be done to eliminate nanoparticles: 1.) Carbon soot needs to be taken out by

filters or change in combustion processes (e. g. HCCI), 2.) Sulphates have to be eliminated as much as

possible by reducing sulphur content in fuel lube or by preventing oxidation of SO2 (e. g. in catalysts),

3.) Eliminating hydrocarbon precursors with oxidation cats (which in turn produce sulphates) or

combustion optimisation (HCCI results in high uHC). Moreover, the fuel quality also plays an

important role: high natural Cetane number and low polyaromatic compounds reduce the number of

carbon particulate precursors.

Indeed, by optimisation of the combustion process from IDI to DI, especially common rail, a

60 % reduction of particulate mass could be achieved. This is due to high fuel injection pressures

over the whole operating range by common rail technology, optimised combustion chamber design,

elaborated liner machining to reduce oil consumption. However, diesel engines still produce a higher

amount of PM mass than SI engines.

DPFs are very effective (up to 95 % removal of PM by mass) if the pore opening is less than

40 to 80 m. Durability of properly operated filter systems has been shown to be exceptional over

as much as 600,000 km. It is shown that increasing the filter diameter/length ratio for any given

filter volume will reduce pressure drop and also decrease peak temperatures in the filter during

uncontrolled regeneration.

Regeneration Strategies

As mentioned above, reducing the balance point temperature is a crucial factor to enable regeneration

at lower temperatures. This can be done by reducing the sulphur content, coating a catalyst such as

platinum onto the wall-flow filter, introducing an oxidation catalyst before the filter (so-called CRT

in which the catalyst oxidises NOx to NO2 which is a strong oxidising agent for soot at 250 C, but

at the same time, it forms sulphate particulates [MAB95]), using non-thermal plasma to replace the

oxidation catalyst (the plasma converts NO to NO2 without significant conversion of SO2 ), or using

fuel-borne catalysts (e. g. iron or cerium additives which strengthen radical-carbon bonds which play

an important role in soot oxidation) [Joh01]. So far, the diesel exhaust temperatures can be too low

to regenerate the filters under all driving conditions, hence, active regeneration techniques are still

needed.

The Peugeot system [SMB00] uses a cerium-based fuel borne catalyst (which reduces the required

regeneration temperature from 550 C to 450 C and in addition increases the soot combustion rate), a

diesel oxidation cat in front of the filter, back pressure monitoring to trigger regeneration (in addition,

the amount of residues from the additive and lube oil ashes are calculated at every moment, as well

2.6. Integrated DeNOx and PM solutions 27

as there is a soot mass estimation based on engine parameter maps; pressure drop measurement is

however necessary after partial regeneration), and engine controls (multiple injection management:

shortened main injection, introduction of a post injection to burn the maximum amount of fuel as

late in the expansion stroke as possible without by-product formation such as CO, uHC, and soot)

to increase exhaust temperature to 450 C but keeping the torque constant, aided by a tuned post

injection (and exothermal reaction in cat) if necessary. The paper does not mention any change

in the EGR control, however, the VGT has to be adjusted due to increased exhaust gas enthalpy

during regeneration. Optimal regeneration frequency (trade-off between increased back pressure fuel

penalty and penalty due to post injection) was established as 400 km. The filter needs to be cleaned

and the additive to be refilled every 80,000 km. The fuel penalty is 5 % for city driving conditions

(less elsewhere) and the maximum torque is reduced by 2 % due to increased back pressure. It is

noted that at speeds higher than 100 kph, the PM number emission becomes larger than without cat

and filter. This is due to sulphate formation on the oxidation catalyst. Hence, a fuel sulphur content

less than 50 ppm is required.

Other regeneration strategies include direct electrical heaters or fuel burners, non-thermal plasma,

and microwave regeneration techniques. None of these appear to be serious contenders at the time

of writing.

Filters combined with SCR seem to have been investigated more than filters combined with traps,

and already are giving promising results. A different approach used a combination NOx trap/filter

consisting of a wall-flow filter with an internal NOx adsorber catalyst coating.

In [KBM+01], a DPF is installed upstream of an SCR catalyst. They report soot removal

efficiencies of 98 % by mass and NOx conversions ranging from 40 to 73 %. A BPT of 315 C at

55 % rated power is achieved using Pt/Ce fuel additives. In their investigations, the NO2 /NOx ratio

over the DPF does not increase with temperature, (NO is catalytically converted to NO2 first but

then used to oxidise soot), therefore a decrease of NOx conversion with decreasing DPF temperature

(which decreases the NOx inlet temperature) is observed.

In [CCH+00], a system combining a NO2 based continuously regenerative trap particulate re-

moval technology with a urea-based Selective Catalytic Reduction (SCR) NOx removal technology

which shows beneficial synergistic effects and leads to simultaneous conversions of 75 to 90 % in

NOx and PM. The feed-gas first passes an oxidation catalyst which converts uHC and CO and in

addition oxidises NO to NO2 . The NO2 subsequently combusts with PM in a particulate filter (CRT).

If some of this NO2 reaches the following SCR catalyst, it will lead to an improvement in the low

temperature NOx conversion of the SCR system (it is shown that a combined CRT + SCR system

results in less NOx than the SCR alone).

In a similar system, Khair et al. [KLF00] put a DPF after an SCR system (hence, avoiding

ammonia slip) and use a cerium fuel additive to facilitate filter regeneration. They achieve 70 % NOx

28 Chapter 2. Diesel Engine Emissions and Aftertreatment

For a successful vehicle application, the integration of NOx traps and DPFs requires further

investigation of the effect of filter regeneration (high temperatures) on the trap and of the effect of

fuel additives to reduce BPT on NOx emission.

Water emulsifed fuel is one of the few primary techniques that can reduce PM and NOx emissions

simultaneously [LPS01]: The break of the correlation between NOx reduction and PM increase

(NOx -PM trade-off) is possible because the presence of water not only physically changes the way

the combustion occurs but also the chemical kinetics of the combustion process. The physical

changes (the reduction of the peak combustion temperature in the presence of water) lead to lower

NOx emissions of up to 19 %, while the chemical changes (different relative quantities of fuel, oxygen,

and inert during the rich premixed-burn stage of diesel combustion, where water acts as a temporary

source of oxygen) lead to reduced PM mass of up to 83 % independent of the temperature change.

The combination of water emulsified fuel in combination with a diesel oxidation catalyst is shown

to reduce PM nearing the level typically offered by a DPF. The big advantage of water emulsified

fuels is that this technology can easily be retrofitted to existing vehicles, therefore speeding up the

process of improving the air quality. However, the maximum water content in diesel fuel is limited

due to a decrease in ignition performance. Moreover, the water has to be prevented from having

contact with metal surfaces in order to avoid corrosion. A new commercially available fuel called

PuriNox achieving this by using an additive which surrounds the water droplets has been certified

by the California Air Resources Board (CARB) in 2001, therefore qualifying for governmental

funding. CARB ascribed the new fuel a 14 % reduction in NOx and 63 % reduction in PM mass

emissions. The costs per kilometer are slightly higher compared to conventional fuel, therefore tax

concessions are needed (and some have already been granted) to encourage the widespread use of

this water emulsified fuel. Note that a break of the NOx -PM trade-off can also be achieved by other

oxygenated agents. In particular, diethylene glycol dimethyl ether which has a high oxygen content

(35.8 mass percent) and high ignitability has been found to simultaneously improve smoke, PM,

NOx , uHC, and noise emissions as well as thermal effiency [MON+98].

2.7 Conclusions

From this survey, it is apparent that no consensus has yet been found on what technology will be

used for diesel aftertreatment. It seems clear, however, that the systems will be complex, and require

very refined engine control to achieve their optimal performance and longevity.

3

Turbocharger Parameterisation for Mean-Value Diesel

Engine Models

The classical purpose of engine models has been to represent the processes in the engine to be able

to simulate its behaviour. Hence, different designs and components can be tested in a computer

environment without the need for expensive prototypes and test cell hours. Moreover, a reliable

simulation model allows for reduction of engine calibration time. With the introduction of modern

control methodologies, models suitable for control design were needed. As opposed to models

for pure simulation purposes, control-orieneted models have to be of limited complexity such that

standard software packets can be used and the designed controller can be implemented in real-time.

The parameterisation of variable geometry turbochargers for mean-value modelling is typically

based on compressor and turbine flow and efficiency maps provided by the supplier. At low tur-

bocharger speeds, and hence low air flows, the heat exchange via the turbocharger housing affects

the temperature-based measurements of the efficiencies. Therefore, the low-speed operating regime

of the turbocharger is excluded from the supplied maps and mean-value models mainly rely on ex-

trapolation into this region. However, operation in this region is commonplace indeed operation

seldom takes place outside this region. Section 3.1 looks at this issue in more detail.

Section 3.2 gives a brief introduction to mean-value modelling of diesel engines with VGT and

EGR. Experimental data for low and medium speed-load points from the turbocharged common rail

diesel engine under investigation are presented in Section 3.3. While the flow maps extend from

the high-speed region in a natural way, the efficiency maps are severely affected by the heat transfer

effect. It is argued that this effect should be included in the mean-value model. Section 3.4 describes

the parameterisation of the turbocharger maps with a focus on the complex turbine efficiency map,

for which a physics-based parameterisation is suggested. This new model structure is then validated

with transient engine data in Section 3.5. Note that for subsequent control design, the transient

performance is more important than the steady-state accuracy, since a reasonable controller will take

care of steady-state errors. A good transient model is required for optimal controller performance.

Both EGR and VGT introduce feedback paths around the cylinders (cf. Figure 3.1). This leads to a

substantial increase in calibration effort. Model-based control aims at reducing this effort but relies

29

30 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

on the underlying model. While crank-angle resolution models are still far too complex for the

purpose of control, mean-value (i. e. cycle-averaged) modelling has been given a lot of attention in

the past. For the airpath, the crucial part of the overall model is the turbocharger. Modelling of this

highly nonlinear device relies typically on performance maps provided by the supplier. However,

due to heat transfer effects and flow measurement problems [MK99], these maps are only provided

for medium and higher turbocharger speeds (i. e. >90,000 rpm for the VGT under investigation in

this project), which are only reached in the higher speed-load range of the engine. In the low and

medium speed-load range, which is important for emission drive cycles for example and where

turbocharger speeds are as low as 10,000 rpm (cf. Figure 3.2 for the New European Drive Cycle

including the extra-urban part), most models simply rely on the extrapolation of the mapped data

[JKS+91, KM95, MK99, MHS98]. In [AFA+01], some lower speed data from a static burner test-

bench are included and parameterised.

Wc i

Ti , p , m i

i

Intake Manifold

Wxi EGR

Tic, p Valve

c

xr

Tr , p

x

Intercooler

Cooler

Tc , p Tx, p , m x

c x

Wxt

Nt

Compressor Turbine

VGT

xv Tt , p t

Ta, pa Oxicat

the turbocharger maps are typically physics-based. A good overview of different parameterisation

methods can be found in [MK99]. The authors particularly address the extrapolation capabilities to

the low speed region, observing some difficulties with neural networks. The elimination of these

problems is reported in [FWA01]. The authors successfully apply an artificial neural network to VGT

modelling. Neural networks can be very helpful for simulation studies; however, they cannot readily

be used for standard model-based control design methodologies. In this project, the turbocharger

maps at low and medium speed-load points are obtained experimentally from a 2.0-litre common rail

diesel engine equipped with a VGT and a turbocharger speed sensor. The data are used to assess the

accuracy of the extrapolation of the supplied maps to lower turbocharger speeds. It turns out that

while the flow maps extend naturally to this region, the efficiency maps are significantly affected by

the heat transfer from the turbine to the compressor side via the turbocharger housing at low flow

3.1. Motivation and Background 31

Compressor Map

3.5

210,000

in

190,000

/p

out

2.5

Pressure Ratio p

170,000

2 150,000

130,000

110,000

1.5

90,000

1

0 50 100 150 200 250 300 350 400 450 500

Compressor Flow in kg/h

Figure 3.2: Typical compressor flow map (blue) and experimental data obtained from a drivecycle

(complete NEDC, red) to point out the importance of the low and medium speed-load region.

rates. The heat transfer decreases the measured temperature before the turbine and increases the

postcompressor temperature, which renders the compressor efficiency artificially low and the turbine

efficiency artificially high. This phenomenon justifies the fact that they are excluded from provided

turbocharger maps. This effect is also observed and implicitly parameterised in [AFA+01] using low

turbocharger speed data obtained from a static burner test-bench, but only for the compressor side.

In mean-value models, the efficiencies are used to calculate the turbocharger speed (via a power

balance between turbine and compressor) and the temperatures postcompressor and postturbine. If

the heat transfer effect is included in both compressor and turbine efficiency, the turbocharger speed

can be predicted correctly. In order to distinguish between aerodynamic efficiency and the efficiency

including heat transfer effects, the latter will subsequently be called pseudo-efficiency. With respect

to the temperatures, these pseudo-efficiencies are actually needed to get the correct values.

While the heat transfer effect on the compressor efficiency does not pose severe problems either

for measurement or for parameterisation, the turbine efficiency is different. At pressure ratios close

to unity across the turbine (which occur at the lower and medium speeds that this study is focused on),

the measured efficiency based on temperatures becomes very sensitive to the pressure fluctuations

of the flow. Moreover, it is difficult to obtain representative temperature readings, because a closed-

coupled oxidation catalyst with its high thermal inertia makes it impractical to wait for the system

to assume equilibrium at each new operating point. At some points, an exothermic reaction in

the catalyst rendered the postturbine temperature even higher than preturbine leading to negative

efficiencies. As will be shown in Section 3.3.4, this problem can be overcome by calculating the

turbine efficiency from the compressor efficiency in steady-state.

The second problem with respect to the turbine efficiency is its parameterisation. This is already

critical without including the effect of heat transfer, since the efficiency depends on the turbocharger

speed, the pressure ratio across the turbine, and the guide vane position. The way forward suggested

in Section 3.4.4 is to separate the heat transfer effect and to parameterise the efficiency map in

32 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

a conventional way using the blade speed ratio [GA98]. The pseudo-efficiency is then obtained

by adding the efficiency due to the heat transfer by modelling the turbocharger housing as a heat

exchanger with flow dependent cooling effectiveness.

The engine model used in this project was developed by Christen et al. [CVC01]. It is event-based

rather than time-based which was shown to be beneficial for applications involving flows [VCG+01],

since the flow-related parameters are less varying in the event domain. The corresponding time-

based model was established in [GA98]. In order to avoid confusion, the equations stated in this

chapter are all time-based, but the conversion to events is straightforward (cf. Section 3.2.2). The

turbocharger submodel in [CVC01] uses a parameterisation of the provided flow and efficiency maps

and extrapolates them into the low speed region. In this study, experimental data is obtained for the

low-speed region, thus making extrapolation unnecessary. Section 3.2.3 familiarises the reader with

the basic equations of the turbocharger modelling using provided flow and efficiency maps. Section

3.2.4 then briefly describes the other subsystems of the engine model, which are joined together to

form the overall model presented in Section 3.2.5. But first, Section 3.2.1 will introduce the notation.

3.2.1 Notation

As convention for the notation in this chapter and throughout the remainder of this dissertation, the

standard symbols for temperature, pressure, etc. contain subscripts referring to the location where

they are measured, e.g. Ti refers to the temperature in the intake manifold. Mass flows are denoted

with two indices, indicating origin and goal. For instance, Wxi is the flow from the exhaust manifold

into the intake manifold, i. e. the EGR flow. An overview of the symbols and subscripts is given at

the beginning of the thesis.

nonlinear differential equation with time as independent variable

= f (x(t), u(t)).

x(t)

This differential equation can be converted to strokes as independent variable by simply applying

the chain rule

= = f (x(), u())

d dt () d e

where the inner derivative gives 1/e since d = dt. Note that the derivative with respect to crank

angle rather than time is denoted by a prime, e. g. e .

3.2. Mean-Value Engine Modelling 33

For the conversion of the engine speed e which is measured in strokes per second, the following

consideration is made: There is one (power) stroke in each cylinder per two (crankshaft) revolutions

in a four-stroke engine, hence, the conversion between [e ] = st/s and [N] = rpm is given as

n cyl N

e =

2 60

where n cyl is the number of cylinders. Flow rates are converted from mass per unit time to mass per

stroke by dividing by the engine speed

W

w= , [w] = kg/st, [W ] = kg/s.

e

3.2.3 Turbocharger

The turbocharger consists of a turbine driven by the exhaust gas and connected via a common shaft

to the compressor, which compresses the air in the intake. The rotational speed of the turbocharger

shaft Nt can be derived as a power balance between the turbine Pt and the compressor side Pc

2

60 Pt Pc

N t = (3.1)

2 Jt Nt

where the turbocharger speed is measured in revolutions per minute (rpm) and Jt is the inertia of

the turbocharger. Subsequently, the expressions for the compressor and turbine power are derived

separately.

Compressor

Assuming that the compression process is isentropic, the following relation between the temperature

and pressure at the inlet (Ta , pa ) and at the outlet (Tc,is , pc ) of the compressor can be derived

1

Tc,is pc

= . (3.2)

Ta pa

However, due to irreversibilities across the compressor (e. g. incidence and friction losses), the

compression process is not isentropic in reality. Therefore, the compressor isentropic efficiency is

introduced which relates the theoretical temperature rise (leading to Tc,is ) to the actual (resulting in

Tc )

Tc,is Ta

c = , 0 < c 1. (3.3)

Tc Ta

Substituting this result into (3.2) yields the expression for the temperature downstream of the com-

pressor:

1

1 pc

Tc = Ta + Ta 1 . (3.4)

c pa

34 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

In order to derive an equation for the compressor power, the first law of thermodynamics is applied

which states that (neglecting heat losses) the compressor power is related to the mass flow through

the compressor Wci and the total change of enthalpy by

where the second equality assumes constant specific heats. Applying (3.4) to this equation finally

gives the expression for the compressor power

1

1 pc

Pc = Wci c p Ta 1 . (3.6)

c pa

In order to calculate the compressor power in (3.6), the compressor efficiency and mass flow have to

be known. These variables are highly nonlinear functions of the pressure ratio across the compressor

and the turbocharger shaft speed. As mentioned in the introduction, these maps are provided by the

supplier for medium and high turbocharger speed obtained from steady-flow test benches. Section

3.3 presents these maps obtained under the pulsating flow conditions in the engine for low and

medium turbocharger speeds.

Turbine

The expressions for the turbine outlet temperature and power can be derived similarly to the com-

pressor outlet temperature (3.4) and power (3.6) yielding

1

pt

Tt = Tx t Tx 1 (3.7)

px

1

pt

Pt = Wxt c p Tx t 1 , (3.8)

px

in which the upper case Pt is the turbine power, while the lower case pt refers to the postturbine

pressure. Again, the turbine flow Wxt and isentropic efficiency t are mapped versus the pressure

ratio across the turbine and the turbocharger shaft speed. However, these variables also depend on the

position of the variable guide vanes, which replace the conventional waste gate to avoid overspeeding

at high engine loads without sacrificing the low load performance. For the turbine flow map, the

dependence on the turbocharger speed can be neglected, however, this is not the case for the turbine

efficiency. Hence, the turbine efficiency map is four dimensional (as opposed to three dimensions

for the other maps), rendering it very difficult to parameterise. Including the effect of heat transfer

complicates the parameterisation even more. However, in Section 3.4, a new physics-based method

to overcome this problem is suggested.

In order to simulate the pressure ratio across the turbine, the exhaust backpressure needs to be

modelled. It is fitted as a quadratic polynomial in the volumetric flow Wxt RTt / pt . This equation

forms an algebraic loop with the turbine equations, and hence, fast dynamics have to be included in

the implementation to break it.

3.2. Mean-Value Engine Modelling 35

This section describes the other subsystems of the engine, i. e. the engine block, manifolds, exhaust

gas recirculation, and EGR- as well as intercooler.

Engine Block

60 Tb Tl

N = (3.9)

2 Je + Jdr

where Je and Jdr are the engine and driveline inertias, respectively. Tl is the load torque and the

brake torque Tb = Tind T f is the difference between the indicated torque (obtained from cylinder

pressure data and approximately proportional to the injected fuel mass per stroke) and the friction

torque (identified from engine data and fitted as a quadratic polynomial in engine speed). The mass

flow rate from the intake manifold into the cylinders is determined by the speed-density equation,

which describes the engine pumping:

m i N Vd

Wie = v (3.10)

Vi 60 2

with the volumetric efficiency v (fitted as polynomial in engine speed and intake manifold pressure)

and the total displacement volume Vd .

For the modelling of the conditions in the exhaust manifold in the next section, the temperature

of the mass flow from the cylinder into the exhaust manifold has to be modelled. This is achieved

by a fitted polynomial in fuel flow, air flow into the cylinders, and engine speed.

Manifolds

The intake and exhaust manifolds are modelled as open thermodynamic systems, where the mass of

gas can increase or decrease with time (so-called filling and emptying model). The two governing

equations for such systems are the Conservation of Mass and the Conservation of Energy. Neglecting

heat losses through the manifold walls and assuming an ideal gas with constant specific heats, the

differential equations for the manifold pressures are derived as

R

p i = (Tic Wci + Tr Wxi Ti Wie ) and

Vi

R

p x = (Te Wex Tx (Wxi + Wxt )) (3.11)

Vx

and for the accumulated masses in the manifolds as

m x = Wex Wxi Wxt . (3.12)

36 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

The manifold temperatures cannot be assumed to be constant and are calculated from the ideal gas

law as

Vi

Ti = pi and

Rm i

Vx

Tx = px . (3.13)

Rm x

Under the assumption that no mass is accumulated in the EGR system it can be modelled with static

equations rather than with differential equations. The flow through the EGR valve is determined by

the standard orifice flow equation [Hey88]

Ar (xr ) px 2 2/ ( +1)/

Wxi = pr pr (3.14)

RTx 1

+1

pi 2

pr = max , (3.15)

px +1

in order to describe subsonic as well as choked EGR flow. The effective area Ar is identified as a

quadratic polynomial of the normalised valve lift xr .

The downstream temperatures of both the EGR- and the intercooler are calculated using the heat

exchanger effectiveness and the appropriate upstream and coolant temperatures

The models of the subsystems from the previous sections can be connected to an overall model

consisting of coupled first-order nonlinear differential equations. The model comprises eight states,

i. e. the intake and exhaust manifold pressures (3.11) and accumulated masses (3.12), the engine

(3.9) as well as the turbocharger speed (3.1), the compressor mass flow (introduced in (3.22) below),

and one state to break an algebraic loop as mentioned in Section 3.2.3. The inputs to the system

are the injected fuel mass, the load torque, and the normalised EGR and VGT actuator positions.

For numerical simulations, the model is implemented in Simulink. The Simulink block diagram is

shown in Figure 3.3. It demonstrates some of the underlying structure and connections between the

subsystems. The main assumptions made to derive this model are:

3.3. Experimental Turbocharger Maps 37

Furthermore, the differences between static and dynamic pressures and temperatures are neglected

due to low gas velocities. These assumptions indicate that the model is of limited complexity

only, which is justified by its purpose for control design. Note that it is not entirely true that the

assumption of no heat loss through the exhaust manifold has been made. In fact, the temperature

rise in the cylinder has been identified from steady-state data of the exhaust gas temperature, thereby

implicitly including heat losses. This appears to be reasonable for the conditions encountered in the

exhaust manifold.

In order to avoid the dependence of the maps on the temperature and pressure upstream of the

compressor and turbine, the following normalised (corrected) quantities are used as is conventional

practice:

T1,up /Tre f

1 = W1 and (3.17)

p1,up / pre f

N t1 = Nt Tre f /T1,up (3.18)

where the subscript 1 refers to the compressor or turbine depending on which map is considered.

The reference temperature and pressure are chosen as 298 K and 101.3 kPa.

Measurement

The compressor flow is measured with a production type mass air flow (MAF) sensor. Although its

accuracy is typically only 7 %, it has the advantage of covering the whole engine operating regime.

Alternatively, the flow can be measured by determining the differential pressure across an orifice

the technique used by the supplier. However, for good accuracy especially at low air flows, a vast

set of orifice plates of appropriate sizes has to be used which was considered impractical for this

application. The postcompressor pressure is obtained from a pressure sensor in the intake manifold

while the pressure upstream of the compressor is assumed to be constant and equal to ambient

conditions. Note that there is a flow dependent pressure drop across the air filter, which can be as

high as 3 kPa at full load. In the operating regime considered here, the maximum pressure drop was

38 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

w_ie

m_i m_i w_ie

w_ci

INTAKE

T_ic T_i T_i w_ex w_ex

(intake manifold)

w_xi p_x

CYLINDERS

p_i p_i T_e T_e

T_r (combustion, mechanics,

emissions) EXHAUST

1 w_f omega 1 w_xi

w_f omega

T_x

T_load T_b

cylinders exhaust

p_i

w_xi

p_x

EGR

T_r

3 x_r

x_r

EGR_system

p_i w_xt

p_x TURBO

w_ci

(VGT, compressor

T_x intercooler)

4 x_v T_ic

x_v

turbo

1 kPa. However, this inlet depression effect is not included in the mean-value model, and hence the

pressure ratio should be measured based on the ambient upstream pressure as well, thereby implicitly

including this effect.

The upstream temperature is also set constant to ambient in the mean-value model. However, on

the engine it increases depending on the operating point by up to 10 K, thereby having a significant

effect on the compressor maps, especially on the compressor efficiency in Figure 3.4 right. As for the

precompressor pressure, rather than modelling this effect it is implicitly included in the compressor

maps by using ambient upstream conditions in the normalisation of the data and the calculation

of the compressor efficiency. This has the advantage of parameterising these effects dependent

on turbocharger speed and pressure ratio without additional modelling effort. The disadvantage is

that the experimental data in Figure 3.4 are not directly comparable to the provided data from the

manufacturer who uses the measured upstream conditions. In order to gain confidence in the accuracy

3.3. Experimental Turbocharger Maps 39

of the measurements, the compressor flow and efficiency have been obtained experimentally using

the measured upstream conditions, which coincided with the provided data quite nicely.

Experimental Results

The experimental data are plotted in Figure 3.4 left. The red crosses denote the data obtained from

the engine for normalised turbocharger speeds in the range from 20,000 to 110,000 rpm, while the

green circles represent the data supplied by the manufacturer. As mentioned earlier, the latter are

only available for speeds larger than 90,000 rpm. The black solid lines correspond to the curve fit,

which is described in Section 3.4.

Compressor Flow Map Compressor Efficiency Map

80

110,000 meas meas

prov prov

1.5 fit fit

70

100,000

Compressor Efficiency in %

60

1.4 110

in

90

/p

out

90,000

50

Pressure Ratio p

1.3

80,000 40

70,000

1.2 30

60,000

20

50,000

1.1 100

90

70 80

40,000 60

10

30,000 40 50

10 20 30

1

0 50 100 150 200 250 300 0

0 50 100 150 200 250 300

Corrected Air Flow in kg/h Corrected Air Flow in kg/h

Figure 3.4: Compressor flow (left) and efficiency (right) map: Measured, provided, and fitted

data for different turbocharger speeds. Note that the provided and measured data are not directly

comparable as discussed in Section 3.3.1. The supplied data are plotted for qualitative comparison

only.

For a fixed turbo speed, the speed lines are very flat at low flows (surge region, inability to

deliver a steady flow against an increasing postcompressor pressure until backflow occurs, which

can seriously damage the compressor) and become very steep for high flows (choke region, the air

velocity reaches the speed of sound and no more flow can be delivered even if the postcompressor

pressure decreases further). The steady-state operating regime of the engine lies between the surge

and choke region as can be seen from the experimental data.

Discussion

The experimental data in Figure 3.4 left confirm what would be expected as a reasonable extrapolation

of the data provided at higher turbocharger speeds. The slight mismatch between the experimental

and supplied data at turbocharger speeds of 90,000 and 110,000 rpm, respectively, can be explained

by the assumption of constant pressure and temperature at the compressor inlet as mentioned in

Section 3.3.1. Using the measured upstream conditions the data coincide very well. It should again

40 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

be pointed out that in the mean-value model, ambient conditions at the compressor inlet are assumed,

and hence, the measured data should be used for parameterisation.

Measurement

The compressor efficiency cannot be measured directly and has to be calculated based on the pres-

sure and temperature ratios across the compressor. Rearranging the expression for the temperature

downstream of the compressor (3.4) yields

1

pc

pa

1

c = (3.19)

Tc

Ta

1

which becomes sensitive to measurement errors at pressure ratios close to unity. However, the flow

on the compressor side is almost steady and the efficiency measurements are stable and repeatable

even for very low turbocharger speeds.

Experimental Results

Figure 3.4 right shows the experimental compressor efficiency data plotted versus the corrected

compressor flow. The efficiency decreases significantly with decreasing corrected turbocharger

speed. This is due to the heat transfer effect mentioned in Section 3.1 (this will become apparent

from the turbine efficiency maps in Section 3.3.4). At low air flows observed at low turbocharger

speeds, the turbocharger housing acts as a quite effective heat exchanger, and transfers heat from

the exhaust to the intake manifold, thereby rising the postcompressor pressure. This results in an

artificially low isentropic efficiency in (3.19), which thus does not reflect the pure aerodynamic

efficiency anymore.

Discussion

The compressor efficiency is used in the mean-value model for the calculation of the intake manifold

temperature and the compressor power. With respect to the temperature, the artificially low (pseudo-)

efficiency is actually needed to predict the correct temperature; otherwise it would be underestimated.

Concerning the turbocharger speed calculation via the power balance (3.1), it will result in the correct

values if the heat transfer effect is also considered on the turbine side, as will be done in Section

3.3.4.

The mismatch between the measured and provided efficiency curves at 90,000 and 110,000 rpm

is significant, but again due to the assumption of ambient conditions at the compressor inlet. As for

the compressor flow, using constant ambient conditions upstream of the compressor in the mean-

value model implies that the measured map should be parameterised, thereby avoiding to model the

pressure drop and temperature rise separately. Although the provided data are therefore not directly

3.3. Experimental Turbocharger Maps 41

comparable to the measurements, they are plotted for a qualitative comparison. Again, using the

measured upstream conditions results in a good agreement with the provided data.

Measurement

While the pressure ratio across the turbine can be measured with the appropriate sensors directly, the

turbine flow needs to be inferred from the mass flow into the engine measured with the MAF sensor.

In steady-state, even with EGR, the turbine mass flow equals compressor mass flow plus fuel flow

(what goes into the engine must come out).

Experimental Results

Figure 3.5 left depicts the experimental data obtained for six different VGT positions. For increasing

VGT position, the turbine restricts the flow more and more, resulting in less flow at the same pressure

ratio. The turbine flow map turns out to be almost independent of the turbocharger speed and therefore

three- rather than four-dimensional.

Turbine Flow Map Turbine Efficiency Map

350

meas VGT = 0%

prov 600 VGT = 40%

0%

fit VGT = 80%

300

20%

500

Corrected Air Flow in kg/h

250

Turbine Efficiency in %

40% 15,000

400

200

60%

300

150

80%

90% 200

100

25,000 55,000

50 100

35,000

45,000

0 0

1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2 1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45

out in in out

Figure 3.5: Turbine flow (left) and efficiency (right) map: Measured, provided, and fitted data for

different VGT positions (0 % is fully open).

Discussion

The measured data extend the provided maps smoothly to pressure ratios down to unity. However,

there are discrepancies at some operating points. As mentioned in Section 3.3.1, these could be

due to measurement errors of the MAF sensor. Therefore, the more accurately obtained provided

turbine flow map has been used for parameterisation at higher pressure ratios, while the measured

data constitute the basis for the extension to pressure ratios close to unity.

42 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

Measurement

As for the compressor efficiency, the turbine efficiency cannot be measured directly, but has to be

calculated from temperature and pressure data. Rearranging (3.7) results in

1 TTxt

t = 1 (3.20)

1 ppxt

which becomes sensitive to measurement errors at pressure and temperature ratios close to one.

As opposed to the compressor side, the pulsating flow in the exhaust causes the pressure ratio

measurement to fluctuate. Moreover, a closed-coupled oxidation catalyst with its high thermal inertia

affects the postturbine temperature. Exothermic reactions in the catalyst can raise this temperature

above the preturbine temperature, leading to negative efficiencies. Besides, it is impractical to wait

for the temperatures to assume equilibrium at each new operating point, especially when going from

high to low loads.

These technical problems can be overcome as follows: According to (3.1), the power balance

between turbine and compressor is in equilibrium once the turbocharger speed (as well as the pressure

ratio and temperatures) have assumed their steady-state values at a given operating point. Hence,

setting (3.1) equal to zero and substituting the expressions for the compressor and turbine power

(3.6) and (3.7), respectively, yields after rearranging

1

Ta

pc

pa

1 W

ci

t c = 1 Wxt (3.21)

Tx

1 pt

px

which is now independent of the postturbine temperature at the expense of the dependence on

compressor efficiency, pressure, and temperature. However, these variables can be measured reliably.

Despite the pulsating flow, steady and repeatable values for the turbine efficiency could be obtained

for turbine pressure ratios down to 1.03. The turbine flow is again simply the sum of the compressor

and fuel flows.

Experimental Results

In Figure 3.5 right, the turbine efficiency is plotted versus the turbine pressure ratio for three different

VGT positions and corrected turbocharger speeds from 15,000 to 55,000 rpm. At pressure ratios

below 1.2, the turbine efficiencies exceed 100 %, which is due to the effect of heat transfer. These

values are calculated based on the compressor efficiency measurements according to (3.21), which

in turn implies that those values are artificially low due to the heat transfer.

For different VGT positions, the speedlines cover a different range of pressure ratios, which is

due to closing the vanes with increasing VGT position and thereby restricting the flow more. As

mentioned earlier, no reliable values could be measured for pressure ratios below 1.03.

3.4. Turbocharger Parameterisation 43

Discussion

At higher turbocharger speeds, the effect of the heat transfer becomes less significant. Ideally, the

measured data would converge to the provided maps, which are available for corrected turbocharger

speeds larger than 50,000 rpm and pressure ratios larger than 1.3. The provided turbine efficiency

values lie in the range between 50 % and 65 % depending on the turbocharger speed and VGT position

(Figure 3.6). Although they are not plotted in Figure 3.5 right for reasons of clarity, the agreement to

the measured values in this range is reasonable. This is especially true considering the observation in

[LGS01] that a loss of up to 30 % in turbine efficiency can be reached under the pulsating conditions

in an engine as opposed to the steady-flow conditions under which the provided maps are obtained.

For the compressor flow map, different parameterisation methods have been developed and tested

in the literature. A good overview can be found in [MK99], which especially investigates the

extrapolation capabilities to lower turbocharger speeds. In this study, the data for the whole low

and medium load operating regime is available and a simple regression leading to a third-order

polynomial in turbocharger speed and compressor flow has been found most straightforward with

satisfying accuracy as can be seen in Figure 3.4 left, where the black lines denote the fitted curves

(the curves are not fitted independently and the emphasis is on the low turbocharger speed region;

the reduced accuracy at higher turbocharger speeds is considered tolerable so that the order of the

fit is not further increased). Additionally, the curve fit was validated in simulations with measured

turbocharger speed as input which decouples the flow from the efficiency maps.

Due to the low slope of the compressor speed lines, it turned out to be beneficial to parameterise

the pressure ratio across the turbine as function of the corrected flow and the turbocharger speed.

However, this introduces an algebraic loop, which can be broken by using the momentum equation

for the mass flow in the tube connecting the compressor outlet and the intake manifold

A

W ci = ( pc pi ) (3.22)

l

where l is the length of the tube and A its cross sectional area. Since the engine is equipped with

the production type intercooler, this pipe is actually more than three meters long for packaging

reasons. Thus, the dynamics are not fast enough to increase the stiffness of the model and thereby

the simulation time significantly.

As for the compressor flow, the efficiency map is parameterised using a regression leading to a

third-order polynomial in turbocharger speed and compressor flow. The fitted curves are depicted

in the right-hand plot of Figure 3.4 (black lines). The regression is reasonably good, especially

44 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

in comparison to the uncertainty that is introduced by the sensitivity of the efficiency to slight

temperature changes.

Due to the nature of the turbine flow, a physics-based parameterisation is the obvious choice using

the standard orifice flow equation (3.14), where the effective area is identified as a quadratic function

of the VGT position and corrected linearly based on the pressure ratio. The correction improves the

fit at low pressure ratios, which are emphasised in this project.

The turbine efficiency map is the most difficult to parameterise. Firstly, this is due to the dependence

of the efficiency on pressure ratio, turbocharger speed, and VGT position rendering the map four-

dimensional. Secondly, the effect of heat transfer increases the range of the efficiencies leading to

pseudo-efficiencies of more than 600 % at pressure ratios close to unity. Regressions have been found

not to be suitable, mainly due to lack of data. The way forward suggested in this study is to separate

the aerodynamic efficiency and the efficiency added due to the heat transfer effect. The aerodynamic

efficiency can then be modelled using a conventional approach based on the so-called blade speed

ratio, while the efficiency due to heat transfer is parameterised based on a heat exchanger equation

with flow-dependent effectiveness. This approach requires the separation of the two efficiencies,

which can be done conveniently by splitting up the postturbine temperature into two terms

This reflects the fact that the measured temperature Tt is equal to the one obtained by using the

purely aerodynamic efficiency Tt,aero in (3.7) minus the temperature drop due to the heat transfer via

the turbocharger housing Tt,heat . Applying (3.23) to (3.20) results in the separation of the pseudo-

efficiency into the aerodynamic efficiency and the efficiency due to heat transfer

Tt,heat

Tx

t, pseudo = t,aero + 1 = t,aero + t,heat . (3.24)

1 pt

px

If no heat loss occurs, the pseudo-efficiency is equal to the aerodynamic efficiency. Even for very

small losses, the efficiency due to heat transfer increases significantly with decreasing pressure ratios

due to their appearance in the denominator. This effect is shown in Figure 3.6, where t,heat is plotted

versus the pressure ratio for Tt,heat being equal to 1 % of Tx . Comparing Figure 3.6 to Figure 3.5

indicates that this might be a reasonable way to parameterise the turbine efficiency.

At low turbocharger speeds, the measured efficiency inevitably includes the heat transfer effect.

Therefore, the supplied data for higher speeds is parameterised and extrapolated to lower speeds.

3.4. Turbocharger Parameterisation 45

800

700

600

500

400

300

200

100

0

1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4 1.45

Pressure Ratio p /p

in out

Figure 3.6: Turbine efficiency due to heat transfer as a function of the turbine pressure ratio for a

1 % loss in temperature.

Potential extrapolation errors will then be compensated in the parameterisation of the efficiency due

to heat transfer.

The conventional approach to parameterise the turbine efficiency is based on the blade speed

ratio defined as

D Nt

cu = (3.25)

1

60 2c p Tx 1 pt

px

where D denotes the turbine blade diameter. Note that this transformation does not introduce new

independent variables. Some of the supplied data are plotted in Figure 3.7, which also shows the

curve fitting.

According to [GA98], the aerodynamic efficiency can be parameterised conveniently as quadratic

function in blade speed ratio

2

cu cu

t,aero = max 2 (3.26)

cu,opt cu,opt

where both max and cu,opt are regressed as polynomials in the VGT position only, without addi-

tional dependence on the turbocharger speed. This can be justified by the fact that the efficiency only

changes little with turbocharger speed and lies well within the accuracy requirements for mean-value

modelling. Moreover, the calculation of the efficiency from heat transfer is based on the measure-

ments and the parameterised values for the aerodynamic efficiency, thereby implicitly compensating

for the modelling errors.

46 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

70

VGT = 0%

VGT = 40%

VGT = 80%

fit

65

Turbine Efficiency in %

60

55

50

45

40

0.4 0.45 0.5 0.55 0.6 0.65 0.7

Blade Speed Ratio

Figure 3.7: Turbine aerodynamic efficiency map: Provided and fitted data for different blade speed

ratios and corrected turbocharger speeds (50, 60, and 70 krpm).

From the parameterisation (3.26), the aerodynamic efficiency can be calculated for a given operating

point. Deducting this value from the measured (pseudo-) efficiency yields the efficiency added due to

heat losses according to (3.24). Mapping these values directly is again a difficult task, since it is not

obvious which independent variables should be chosen to avoid parameterising a four-dimensional

map (i. e. based on corrected turbocharger speed, pressure ratio, and VGT position). It is therefore

suggested to employ the physics behind the heat transfer process, and hence to model the turbocharger

as a heat exchanger. As will be seen later, this reduces the map to three dimensions and the data

confirm that this assumption is indeed reasonable.

The governing equation for a heat exchanger has already been introduced in (3.16). For the

turbocharger, the upstream temperature is the exhaust manifold temperature, and the coolant tem-

perature is chosen as the intake manifold temperature. From (3.16), the temperature drop due to heat

transfer via the turbocharger housing can be derived as

where V GT describes the cooling effectiveness of the turbocharger. For each engine operating point,

the cooling effectiveness can be calculated by first deriving Tt,heat from t,heat in (3.24) and then

applying (3.27). The effectiveness should decrease with increasing flow rate through the turbine

which is confirmed in Figure 3.8, where V GT is plotted versus the corrected turbine flow. Hence,

the pseudo-efficiency converges to the aerodynamic efficiency at larger flows. The dependence on

the VGT position is also reasonable, since the flow velocity has to be higher to achieve the same

flow rate through a more restricted turbine. This implies that the effectiveness for the same flow rate

should be less if the VGT is more closed which is confirmed in Figure 3.8.

3.5. Experimental Validation 47

30

VGT = 0%

VGT = 40%

VGT = 80%

fit

25

Cooling Effectiveness in %

20

15

10

0

20 40 60 80 100 120 140 160 180 200

Figure 3.8: Cooling efficiency of the turbocharger for different corrected turbine flows and VGT

positions.

Figure 3.8 also shows the curve fit, which has been chosen as

1

V GT = 30 exp (x x0 ) + x1 (3.28)

15

In models for control, the steady-state accuracy is less important since any reasonable controller

will take care of steady-state errors. It is rather the dynamic behaviour of the plant especially at

frequencies around the intended bandwidth of the closed-loop system, which is important. While

this section presents transient simulation and experimental data for time-domain validation, the

frequency response of the system will be identified and compared to the model response in the next

chapter, which focuses on the control design in frequency domain.

In this project, the focus of the modelling and validation is on the turbocharger, which is the crucial

nonlinear part of the overall engine model. In order to avoid that errors in other submodels propagate

to the VGT model, the crankshaft dynamics described in (3.9) are bypassed and the measured engine

speed is used instead. Moreover, the EGR valve is kept shut to decouple the EGR from the VGT

system. Note that the frequency domain validation in the next chapter includes nonzero EGR valve

positions.

In order to validate the model over a range of engine operating points, 50 seconds of the extra

urban part of the New European Drive Cycle (NEDC) are chosen as transient excitation. The engine

speed, fuel rate, and VGT position are used as input to the simulation model. A comparison of the

experimental and simulation data is given in Figure 3.9. The overall agreement is quite good, but

overshoots in the simulated turbocharger speed result in overshoots in the intake manifold pressure,

48 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

2500 150

2000 100

N in rpm

Tl in Nm

1500 50

1000 0

0 10 20 30 40 50 0 10 20 30 40 50

140 200

sim

130 exp

MAF in kg/h

MAP in kPa

150

120

100

110

100 50

0 10 20 30 40 50 0 10 20 30 40 50

100

100

VGT Pos in %

N in krpm

50

50

t

0 0

0 10 20 30 40 50 0 10 20 30 40 50

80 150

60

eta in %

etat in %

100 pseudo

40

c

50 aero

20

h.t.

0 0

0 10 20 30 40 50 0 10 20 30 40 50

exp

pratc

pratt

1.2 1.2

1 1

0 10 20 30 40 50 0 10 20 30 40 50

310 800

700

300 600

Tx in K

Ti in K

500

290

400

300

280

0 10 20 30 40 50 0 10 20 30 40 50

t in sec t in sec

Figure 3.9: Experimental and simulation results for part of the New European Drive Cycle (NEDC);

red: experimental data, blue: simulation data, black: if used for both.

3.5. Experimental Validation 49

and hence in the pressure ratio across the compressor. However, this can be improved by either

increasing the turbochargers inertia or by low-pass filtering the VGT position input.

It could be argued that the over- and undershoots in the turbocharger speed are due to the fact

that the effect of heat transfer on the efficiencies is modelled statically although it might be a rather

slow process compared to the transients encountered in the NEDC. However, it is the aerodynamic

efficiency, which determines the turbocharger speed on the engine. The heat transfer effect is only

included in the model because it cannot be separated when inferring the turbocharger efficiencies

based on temperature measurements and it is needed to give the correct postcompressor and turbine

temperatures. Hence, if the measured compressor and turbine efficiency maps, which include the

effect of heat transfer, are to be used rather than the extrapolation of the aerodynamic efficiencies, the

artificially low compressor efficiency will be compensated by the artificially high turbine efficiency,

thereby cancelling out the heat transfer effect when determining the turbocharger speed. Figure 3.9

shows that this works well for turbocharger speeds as low as 40,000 rpm. At even lower speeds, the

pressure ratios across the compressor and turbine are close to unity resulting in a high sensitivity to

modelling errors (cf. Figure 3.5 right). Therefore, a lower limit of 10 % for the compressor efficiency

had to be chosen to avoid stalling of the turbocharger at very low pressure ratios in the simulation.

Note that the turbocharger speed between 30,000 to 90,000 rpm lies mostly well outside the range

of the provided maps. The effect of heat transfer is significant, which can be seen in the compressor

and turbine efficiency plots. The turbine pseudo-efficiency exceeds 100 % and the efficiency due to

heat transfer (magenta line), which is added to the aerodynamic efficiency (black line), is significant.

This indicates that the concept of including the heat transfer effect in order to use the experimentally

obtained efficiency maps works well.

135 200

sim

exp

130

125

150

MAF in kg/h

MAP in kPa

120

115

100

110

105

100 50

0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50

t in sec t in sec

Figure 3.10: Experimental (red) and simulation (blue) results for MAP and MAF with measured

turbocharger speed as model input.

With respect to the intake and exhaust manifold temperatures, the lowest plots in Figure 3.9

show some mismatches especially during transients. This is partly due to the low-pass effect of the

thermocouples. For the exhaust manifold, differences in steady-state are due to inaccuracies in the

50 Chapter 3. Turbocharger Parameterisation for Mean-Value Diesel Engine Models

parameterisation of the exhaust temperature which only implicitly includes heat transfer effects as

mentioned in Section 3.2.5.

The compressor and turbine flow maps can be validated separately from the efficiency maps by

feeding the measured turbocharger speed into the model. The results for this setup are shown in

Figure 3.10. While the manifold air flow is slightly underestimated, the manifold absolute pressure

coincides with the measured data very well. Note that the MAP is the output of the parameterised

compressor map based on turbocharger speed and MAF. Due to the flat compressor lines in the

operating region (cf. Figure 3.4 left), the pressure ratio is not very sensitive to errors in MAF. The

equilibrium MAF and MAP levels are determined from an equilibrium of the engine pumping rate

(3.10) and the compressor flow map. The mismatch in MAF in Figure 3.10 is therefore probably

due to an underestimation of the volumetric efficiency v at that operating point.

Further validation will be undertaken in the next chapter. It also includes an evaluation of the

effect of parameter variations on the frequency response of the system.

4

Airpath Model and Plant Investigation

The first section of this chapter will give an overview of the issues associated with the EGR-VGT

control problem. Since the control design tools to be applied to this problem in Chapter 5 require

a linear model of the plant, the effects of the linearisation around a nominal operating point are

investigated in Section 4.1.1 by comparing the nonlinear to the linearised model. Furthermore, this

investigation is used in Section 4.1.2 to familiarise the reader with the properties of the EGR-VGT

system.

The control design methodologies used in this dissertation are all frequency domain techniques.

This is the motivation to compare the frequency response obtained from the linearised model to the

experimentally identified frequency response for further validation in Section 4.2. This procedure

also gives deeper insight into the physical properties of the plant. In order to explain observed

differences between the simulation and experimental data, different parameters of the engine model

at hand are varied and their effects on the frequency response are evaluated in Section 4.3.

As already mentioned in Section 1.1.1, EGR and VGT are both driven by the exhaust gas and therefore

constitute a multivariable system. The corresponding model is obtained by rearranging the model

derived in the previous chapter. As depicted in Figure 4.1, the inputs are the normalised positions

of the EGR (xr in %) and the VGT (xv in %) actuator. The outputs of the system are the mass air

flow into the intake manifold MAF (Wci in kg/h) and the intake manifold pressure MAP ( pi in kPa).

Engine speed (N in rpm) and fuelling rate (W f in mg/st) are external inputs, which can be regarded

as parameters.

The reason for using MAF and MAP as model outputs is that the engine is equipped with sensors

for these signals. Alternatively, references for internal variables (e. g. the burnt-gas fraction in the

intake manifold for the EGR control, which could be easily introduced to the model) can be tracked

by actuating EGR and VGT as well, if there is an estimator for theses quantities available. However,

the investigations in [NKM00] lead to the conclusion that the choice of setpoints is more crucial

than the strategy to achieve them. In this study, only MAF and MAP feedback will be investigated.

This has the advantage that the reference values from the engine strategy can be used directly, which

are optimised in steady-state with respect to emissions and are mapped over the entire speed and

fuelling envelope.

51

52 Chapter 4. Airpath Model and Plant Investigation

Over most of the operating envelope, MAF is mainly driven by EGR and MAP by VGT. Therefore,

production type engine controllers typically close both loops separately with PI(D) controllers, which

have to be detuned to allow for the ignored cross-couplings. In transients, the control objective is

to provide sufficient fresh air charge at pedal tip-in so that enough fuel can be burnt without visible

smoke to give the demanded torque, thereby avoiding the so-called turbo-lag. Moreover, sufficient

recirculated exhaust gas is needed to keep the NOx emissions low.

Engine

VGT position MAP

The control design procedures applied in Chapter 5 are model-based and require linear time-invariant

models. It is therefore necessary to linearise the nonlinear engine model from Section 3.2.5 around a

chosen operating point. This linearisation is performed algebraically from the nonlinear differential

equations using the Symbolic Toolbox of Matlab. The nominal operating point has been chosen

to lie in the medium speed and load region at 1500 rpm, 85 Nm. The conditions at this operating

point (in nonlinear simulation as well as in experiment) are given in Table 4.1. The EGR and VGT

actuator positions correspond to the measured values when running the supplied engine strategy at

the nominal operating point. Since the exhaust back pressure does not change significantly around a

fixed speedload point, it has been chosen as a constant (obtained from nonlinear simulation) in the

linearisation, thereby avoiding an additional state to break the algebraic loop as mentioned in Section

3.2.3. Moreover, due to the linearisation for a fixed engine speed (i. e. 1500 rpm), the crankshaft

dynamics in (3.9) can be bypassed so that the engine speed becomes an external input of the model

as mentioned above rather than a state. Hence, the linearised model is of sixth order, where the states

correspond to the internal variables and outputs listed in Table 4.1.

The state-space matrices are provided here as well so that the interested reader can follow the

control design in the subsequent chapter:

0.4125 0.0248 0.0741 0.0089 0 0

101.5873 7.2651 2.7608 2.8068 0 0

0.0741 0.0089 0.0200

0.0704 0.0085 0

A= ,

0.0878 0.2672 0 0.3674 0.0044 0.3962

1.8414 0.0343 0.0330

0.0990 0 0

0 0 0 359.0000 187.5364 87.0316

4.1. The EGR-VGT Control Problem 53

0.0042 0.0064

1.0360 1.5849

0.0042

0 0 0 0 0 0 3.6 0 0

B= , C = , D= .

0.1261 0 0 0 0 1 0 0 0 0

0.0168

0

0 0

In order to gain confidence in the linearised model, the responses to step changes around the nominal

values of EGR (xr ) and VGT (xv ) are compared in simulation for the linear and nonlinear model in

Figure 4.2. On the engine under investigation, EGR and VGT are both driven by electronic vacuum

regulator valves (EVRV), which adjust the vacuum pressure delivered to the actuators according

to a electric duty cycle signal generated by the engine control unit (ECU). In order to be able to

compare the simulation data to experimental data in the next section, a step change in the duty cycle

is commanded to the EVRV of the EGR and VGT, respectively (cf. upper plots in Figure 4.2). The

resulting actuator position response is recorded (red lines) and then used as input to the linear and

the nonlinear model.

Table 4.1: Conditions at the nominal operating point in experiment and simulation.

Inputs

Engine Speed [N] = rpm 1500 1500

Fuelling [W f ] = mg/st 16.7 16.7

EGR actuator position [xr ] = % 20 20

VGT actuator position [xv ] = % 60 60

Outputs

Flow into intake manifold (MAF) [Wci ] = kg/h 78.5 84 x6

Intake manifold pressure (MAP) [ pi ] = kPa 111.3 111.0 x4

Internal Variables

Intake manifold mass [mi ] = g 7.15 7.18 x3

Exhaust manifold pressure [ px ] = kPa 116.5 118.0 x2

Exhaust manifold mass [mx ] = g 0.66 0.64 x1

Turbocharger speed [Nt ] = rpm 58,400 61,000 x5

Figure 4.2 shows the MAF and MAP response of the system, where the simulation results based

on the linear model are plotted in green and the blue lines correspond to the nonlinear model. These

data coincide very well except for a slight steady-state offset in the MAF response to the EGR

input, and hence give the confidence that the linear model represents the nonlinear model at the

chosen operating point rather well. Note that the red lines represent the data obtained from engine

experiments, which will be discussed in the next section.

54 Chapter 4. Airpath Model and Plant Investigation

Actuator in % 100 100

50 50

position

duty cycle

0 0

0 5 10 15 20 25 0 5 10 15 20 25

MAF in kg/h

nl

lin

90 90

80 80

70 70

0 5 10 15 20 25 0 5 10 15 20 25

120 120

MAP in kPa

115 115

110 110

105 105

0 5 10 15 20 25 0 5 10 15 20 25

t in sec t in sec

Figure 4.2: Comparison of the MAF and MAP response to step inputs in EGR and VGT; experimental

data (red), nonlinear model (blue), linear model (green).

Actuators

Due to their operating principle, the aforementioned EVRVs introduce further nonlinearities to the

system, which cannot be neglected. These nonlinearities include (cf. Figure 4.2): static gain change

depending on the applied duty cycle, saturation, deadband, different dynamic response depending on

the amplitude of the step change (faster response for larger steps), different time constants depending

on the direction of the applied step input (due to working against or with the spring in the actuator

and faster venting of the vacuum than building it up), and hysteresis (cf. the VGT position in the top

right plot; although the duty cycle is stepped down to its original value of 72 %, the position only

goes down to 55 % rather than to its original value of 50 %).

Modelling of the nonlinear behaviour of an EVRV is a substantial task; the interested reader is

referred to [MKN99]. However, both the EGR and VGT actuator are equipped with position sensors

so that position controllers (PI type) can be implemented. These position control loops will be

identified as first order systems in Section 4.2 and then be included in the overall model for control

design.

4.1. The EGR-VGT Control Problem 55

Engine

The discussion of the plant properties in this section is mainly based on the open-loop MAF and

MAP responses to steps in the EGR and VGT command at the nominal operating point in Figure

4.2. Subsequently, each channel will be considered separately and the occurring phenomena will be

explained based on the underlying physics. A further investigation of the plant behaviour over the

speed-load envelope can be found in [KMN+97] and [MNK97]. However, the results in there are

solely based on simulations.

Increasing the EGR position corresponds to opening the EGR valve and thereby increasing the EGR

flow, and hence reducing the MAF into the engine (cf. middle left plot in Figure 4.2). Note that

the response in this channel is almost instantaneously, which justifies the fact that the EGR system

is modelled statically. Comparing the experimental data to the simulation results shows very good

agreement for the dynamic response and for the gain of the system. The steady-state offset is not

of concern and will be corrected by any viable controller. The investigations in [KMN+97] show

that the EGR valve loses authority over the MAF at low engine speeds, suggesting that the VGT

would be more suitable to track MAF under those conditions. This can be used in a multivariable

control scheme to increase the speed of response at low engine speeds and loads by using both the

EGR and VGT to give a fast MAF response and therefore small turbo-lag. In fact, investigations in

[NMK+98] show that this is the only part of the engine envelope, where a coordinated EGR-VGT

control achieved better performance than a decentralised strategy.

Increasing the VGT position corresponds to closing the guide vanes such that the exhaust gas flow is

more restricted, which increases the turbocharger speed and results in more compression of the inlet

charge, i. e. increased MAP. When the guide vanes are opened, the VGT acts like a wastegate such

that turbocharger speed and hence MAP decrease. Again, a comparison between experimental and

simulation data in Figure 4.2 (bottom right plot) reveals a steady-state offset which is not of concern.

While the dynamics of the MAP response to an increase in the VGT position are captured very well,

the simulated response (also with the nonlinear model) to a decrease is faster than actually observed

on the engine. This behaviour is not due to the actuator position control as can be seen in the VGT

position trace in the upper right plot. However, a look at the turbocharger speed trace (not plotted

here) reveals that it is the turbocharger which slows down so slowly. Further investigation shows

that the exhaust manifold pressure builds up quite quickly once the guide vanes are closed, but it

takes some time to release the pressure once the guide vanes are opened again.

Again, increasing the EGR position setpoint is equivalent to opening the EGR valve, which means

that more exhaust gas is recirculated into the intake manifold resulting in a slight increase in MAP.

56 Chapter 4. Airpath Model and Plant Investigation

However, due to increased EGR, less exhaust gas enters the turbine, thereby slowing down the

turbocharger eventually, and hence decreasing MAP. The latter effect is dominant but slower due

to the turbocharger dynamics. This explains the non-minimum phase behaviour in the bottom left

plot in Figure 4.2. Note that the non-minimum phase behaviour is very well captured by the model

resulting in a right half-plane zero. However, the system gain is underestimated, the actual effect of

EGR on MAP is stronger than in the model. In fact, it is as significant as the effect of the VGT at this

operating point. At higher speed and load points, the EGR to MAP coupling is weaker than the EGR

to MAF coupling though. As mentioned before, this finding is confirmed in [KMN+97], where a

gain array analysis suggests that in a decentralised strategy MAP should be driven by EGR and MAF

by VGT at low speed and load points. The chosen operating point for this study (1500 rpm, 85 Nm)

seems to lie somewhere inbetween. The reason for this gain underestimation will be addressed in

Section 4.3 again.

Increasing the VGT position, i. e. closing the guide vanes, leads to a greater restriction of the exhaust

gas flow and increase in exhaust manifold pressure. This results in an increased EGR flow which

reduces MAF as can be seen in the the middle right plot in Figure 4.2. However, if the EGR valve

is completely closed, the EGR flow cannot be increased, all of the exhaust gas must pass over the

turbine, therefore spinning it up if the guide vanes are closed. This leads to an increase in compressor

power, and hence in MAF and MAP as shown in Figure 4.3. Therefore, the steady-state gain from

VGT to MAF undergoes a sign change depending on the operating point. Since it is uncertain where

the sign change takes place, the VGT should not be used to track MAF setpoints in a decentralised

strategy (e. g. integral SISO control requires the sign of the static gain of the plant to be known, while

integral MIMO control only imposes constraints on the sign of the determinant of the static plant

gain matrix, hence elements of it might well change their signs [Lun97]).

Comparing simulation and experimental data in Figure 4.2 as well as in Figure 4.3 shows that the

gain of the system is correctly modelled, but for the same reasons as for the VGT to MAP channel,

the response time for opening the VGT is underestimated. Moreover, Figure 4.2 shows a small

overshoot in the simulation data which does not occur in the experiment. Note that Figure 4.3 does

not show the data from the linear model because the equilibrium point is different (EGR is now 0 %

rather than 20 %).

This section describes the experimental validation of the frequency response at the nominal operating

point (1500 rpm, 85 Nm) in order to further validate the model derived in Chapter 3. The sampling

frequency has been chosen as 125 Hz and is therefore larger than twice the engine event frequency

(50 Hz at 1500 rpm) to avoid aliasing.

The spectral data of the inputs and outputs of the system (such as in Figure 4.4 on the right side)

indicate that the system behaves approximately linearly for small perturbations around the nominal

4.2. Frequency Domain Validation 57

100

Actuator in %

50

0

0 5 10 15 20 25

120

MAF in kg/h

115

110

105

100

0 5 10 15 20 25

130

exp

MAP in kPa

125 nl

120

115

110

0 5 10 15 20 25

t in sec

Figure 4.3: Comparison of the MAF and MAP response to a step input in VGT for closed EGR;

experimental data (red), nonlinear model (blue).

operating point because the output spectrum is mostly limited to the frequency range of the input

spectra. This holds also true for higher frequencies than the ones plotted in the figure and confirms

the findings from Section 4.1.2, where the linear model was shown to represent the plant behaviour

quite well (except for steady-state offsets and the overestimation of the time constant for the MAF

and MAP response when the VGT is opened). Therefore, the concept of frequency responses, which

is only defined for linear systems, can be applied to the validation problem at hand. Note that small

peaks appear at multiples of the cylinder event frequency (corresponding to 12.5 Hz at 1500 rpm)

even for fixed EGR and VGT positions.

4.2.1 Procedure

In order to experimentally obtain the frequency response of the engine from the EGR and VGT

inputs to the MAF and MAP outputs, respectively, the system is excited with sinusoids of different

frequencies in one channel, while the other actuator is fixed in a constant position. Rather than

conducting consecutive experiments with sinusoids of different frequencies, a summation of such

sinusoids is chosen as input signal. Therefore, the frequency range of interest is divided into equidis-

tant frequencies, and sinusoids with same amplitude at these frequencies are added. The phase of

these sinusoids is also distributed in order to avoid large peaks in the resulting signal. The period

of the generated signal is equal to the period of the lowest frequency chosen. Figure 4.4 shows for

the EGR to MAF channel one period of the resulting signal for a frequency range from 0.01 Hz to

1 Hz (upper left plot) and the corresponding Fast Fourier Transform (FFT) (upper right plot). Note

that the power spectrum of the signal is equally distributed over the frequency range. The transfer

function of the system can then be obtained directly by dividing the output by the input magnitude

spectrum (the phase is obtained by deducting the output phase from the input phase, of course).

58 Chapter 4. Airpath Model and Plant Investigation

[ EGR Ref] = % 10 60

40

0

20

10

0

0 20 40 60 80 100 0 0.5 1 1.5 2

[ EGR Pos] = %

10 60

40

0

20

10

0

0 20 40 60 80 100 0 0.5 1 1.5 2

10 60

[ MAF] = kg/h

40

0

20

10

0

0 20 40 60 80 100 0 0.5 1 1.5 2

t in sec f in Hz

Figure 4.4: Frequency domain validation methodology for the EGR to MAF channel: Time signals

on the left, frequency contents on the right side. Steady-state offsets are removed.

Due to the dynamics of the actuator position control loops, the reference positions of EGR and VGT

are low-pass filtered. This implies that the actuators ability to track a reference signal becomes less

with increasing frequency. This can be seen in Figure 4.4 in the plots of the middle row, which show

the time signal and spectrum of the actual position as opposed to the reference position shown in

the plots of the upper row. From this data, the transfer function of the position control loop for the

EGR valve is obtained and plotted in Figure 4.5 (left). The plots on the right-hand side of the same

figure show the transfer function of the VGT actuator position control loop obtained from a separate

experiment.

Although the EVRVs and the PI control for the EGR and VGT actuation are identical, the closed-

loop bandwidth (-3 dB) is larger for the EGR than for the VGT system (approximately 1.0 Hz as

opposed to 0.5 Hz). This is due to the different designs of the valves (e. g. smaller stroke of the

EGR valve). Figure 4.5 also shows first order approximations of the position control loops, which

need to be included in the overall model for subsequent control design. Comparing the experimental

and modelled phase plots shows that there is no significant additional phase lag in the frequency

range of interest. Hence, time delays due to CAN bus interfacing (for this application up to 50 ms,

cf. [YC01]) and due to transport delays in the vacuum delivery to the EGR and VGT actuators can

be neglected.

4.2. Frequency Domain Validation 59

EGR VGT

0 0

10 10

Gain

Gain

exp

model

1 1

10 10

2 1 0 2 1 0

10 10 10 10 10 10

0 0

20 20

Phase

Phase

40 40

60 60

80 80

2 1 0 2 1 0

10 10 10 10 10 10

f in Hz f in Hz

Figure 4.5: Identified frequency responses of the actuator position control loops.

The closed-loop transfer functions for the EGR and VGT actuator position control loops plotted

in Figure 4.5 are chosen to

2 1.0 2 0.5 Pactegr

Pactegr = , Pactvgt = ; Pact = (4.1)

s + 2 1.0 s + 2 0.5 Pactvgt

Note that these bandwidths correspond quite well to the open-loop bandwidths of the actuators that

can be estimated from the time constants obtained from the step responses in Figure 4.2.

The transfer functions from the plant inputs (EGR and VGT actuator position) to the outputs (MAF

and MAP) are obtained from the FFT of the actual actuator positions (i. e. the middle plots in Figure

4.4 for the EGR actuator) and the FFT of the measured output signals (i. e. the bottom plots in Figure

4.4 for the MAF signal) as described in the previous section for the position control loops. The sum

of sinusoids input is applied to each actuator in a separate experiment, while recording both MAF

and MAP simultaneously. With two inputs and two outputs, the system is characterised by a 2 2

transfer function matrix as shown in Figure 4.6. Note that the outputs have been scaled to make the

frequency responses comparable to each other (cf. Section 5.1.1).

Figure 4.6 also shows the frequency response of the linearised model at the nominal operating

point. Since the model has been parameterised with steady-state experimental data, one would expect

good agreement at least for low frequencies. This is confirmed by the experimental data for the main

loops, i. e. EGR to MAF (upper left plot) and VGT to MAP (lower right plot), although the gain is

overestimated at higher frequencies. Due to the limited bandwidth of the actuator position control

loops as described in the previous section, the maximal frequency for the validation has been chosen

to 1 Hz, which is approximately the bandwidth of the EGR position control loop; higher frequencies

60 Chapter 4. Airpath Model and Plant Investigation

EGR VGT

1 1

10 10

0 0

10 10

MAF

1 1

10 10

2 2

10 10

2 1 0 1 2 2 1 0 1 2

10 10 10 10 10 10 10 10 10 10

1 1

10 10

model

exp

0 0

10 10

MAP

1 1

10 10

2 2

10 10

2 1 0 1 2 2 1 0 1 2

10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz

Figure 4.6: Frequency response of the scaled plant at 1500 rpm and 85 Nm.

cannot be tracked properly by the actuators anymore. Note that the target bandwidth of the overall

closed-loop system has to be significantly smaller than the underlying actuator position control loop

to avoid undesired interactions.

The cross-couplings between the EGR and VGT system are characterised by the off-diagonal

plots in Figure 4.6. As opposed to the main loops, the low frequency experimental data differ

significantly from the model data. The low frequency gain from EGR to MAP is underestimated by

a factor of almost three indicating a problem with the model, which will be further investigated in

the next section. This underestimation of the gain can also be seen in the bottom left plot in Figure

4.2, where the change in MAP is significantly smaller in simulation than in experiment.

With respect to the cross-coupling between VGT and MAF, the low frequency gain is also

underestimated. However, this is not confirmed by the step response in the middle right plot in Figure

4.2. By repeating the frequency validation procedure, it has been found that all data are coinciding

very well except for some data scattering in the low frequency of the VGT to MAF channel. The

mean of that scatter is close to the frequency response of the model so that the observed difference

between experimental and simulation data in the upper right plot of Figure 4.6 is not consistent. The

reason for this behaviour is the sensitivity of the low frequency gain of the VGT to MAF channel

with respect to slight changes in the EGR valve position. This becomes clear by recalling the change

4.3. Model Uncertainty Investigation 61

of gain observed between Figure 4.2 and Figure 4.3, where also the sign changes when the EGR

position is reduced from 20 % to 0 %. In the experiment, the EGR position setpoint is fixed at

20 %, but exciting the VGT leads to changes in the exhaust manifold conditions which constitute a

disturbance for the EGR valve position such that it actually changes by 1 % (in absolute position).

This seems to be sufficient to cause the observed scatter in the low frequency data of this channel.

Finally, the slopes of the frequency responses in the VGT to MAF and VGT to MAP channel do

not match well at frequencies larger than 0.2 Hz. This indicates missing poles in the model and is

potentially due to some neglected dynamics in the coupling from the VGT position via the exhaust

manifold and EGR system to the MAP and MAF in the intake manifold.

In order to better understand the differences between the modelled and the identified frequency

response in Figure 4.6, the sensitivity of the model to parameter changes is investigated. This is

done by increasing and decreasing the parameter values in the nonlinear model by 15 %. It turns

out that most of the parameters only have a minor effect on the frequency response. However, three

parameters (the turbine and compressor efficiencies as well as the effective area of the EGR valve) do

have a significant effect. Moreover, this effect turns out to be exactly where the differences between

simulation and experiment are observed, namely in the low frequency range of the cross-couplings.

While the efficiencies mainly affect the EGR to MAP channel, the effective area of the EGR valve

influences the VGT to MAF channel. Figure 4.7 shows the nominal frequency response (black

lines) along with the boundaries that are obtained from the worst combination of the three parameter

changes (red lines).

Note that the parameters identified to have the biggest effect are the ones that one would suspect to

be most uncertain. The modelling in the previous chapter has shown that the turbocharger efficiencies

are most difficult to parameterise, and the uncertainty in the effective area of the EGR valve stems

from the fact that it has been identified from averaged experimental data and parameterised as

nonlinear function of the valve lift. Moreover, the orifice model for the EGR valve is only a crude

approximation of the true phenomena; all the unmodelled effects are lumped into the effective area.

Both the efficiencies and the effective area cannot be measured directly.

The sensitivity of the frequency response in the VGT to MAF channel with respect to the effective

area of the EGR valve has already been pointed out in the discussion of this channel in Section 4.1.2.

Hence, it remains to discuss the effect of the turbocharger efficiencies on the EGR to MAP channel.

Note that both the compressor and the turbine efficiency have similar effects on the frequency

response, which leads to the conclusion that their effect on the turbocharger speed is crucial. In the

bottom left plot in Figure 4.7, the red line close to the experimental data is obtained by increasing all

parameters by 15 % (where the change in the effective area has only a minor effect). An increase in the

efficiencies corresponds to a higher turbocharger speed; this indicates that the model underestimates

the turbocharger speed, which is confirmed by comparing the experimental and simulated speeds in

Table 4.1. The table also reveals that the steady-state MAP values coincide pretty well, while steady-

state MAF is also underestimated in simulation. Now, consider the compressor flow map in the left

62 Chapter 4. Airpath Model and Plant Investigation

1

EGR 1 VGT

10 10

0 0

10 10

MAF

nominal

1 1

10 limits 10

exp

2 2

10 10

2 1 0 1 2 2 1 0 1 2

10 10 10 10 10 10 10 10 10 10

1 1

10 10

0 0

10 10

MAP

1 1

10 10

2 2

10 10

2 1 0 1 2 2 1 0 1 2

10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz

Figure 4.7: Frequency response of the scaled plant at 1500 rpm and 85 Nm and with varied param-

eters (combination of 15 % changes in c , t , and Ar ).

plot of Figure 3.4 and remember that a change in EGR position firstly affects MAF. By plotting the

equilibrium points from the simulation (with the underestimated turbocharger efficiencies) and the

experiment on the compressor flow map and considering oscillations around the equilibrium MAF,

one can see that the effect on the MAP is stronger in the experimental case where the tangent to the

flow lines is steeper. This is exactly what is seen in Figure 4.7.

In the next chapter, the effect of different parameterisations of the observed uncertainty on the

control design will be investigated. These uncertainty descriptions will be based on the bounds on

the frequency response as shown in Figure 4.7.

5

Uncertainty Parameterisations and Robust Control Design

Based on the mean-value model for the airpath of the engine developed in Chapter 3, robust multi-

variable controllers will now be designed and compared which actuate the EGR valve and the VGT

guide vanes in order to track a commanded MAF and MAP signal. The applied design methodolo-

gies, namely H loopshaping and synthesis, are described in Appendix A, which also gives an

introduction to robust control theory.

The essential feature of robust control is to explicitly account for uncertainties in the model

of the plant. While an H loopshaping controller that assumes a general type of uncertainty (so-

called coprime factor uncertainty) is designed and tested in Section 5.1, its performance will be

improved by an extended loopshaping procedure resulting in a two degrees of freedom controller

for coprime uncertainty (CUC) in Section 5.2. The potential improvements that can be achieved

by parameterising the identified uncertainty specifically are investigated in Section 5.3, where an

additive uncertainty controller (AUC) is designed.

Further uncertainty parameterisations are obtained in Section 5.4 and a state-space approach

turns out to give a good description of the observed uncertainties yielding real uncertain parameters

as opposed to the complex parameters in the coprime and additive uncertainties. The previously

designed CUC and AUC are then compared in Section 5.5; firstly using mixed analysis based

on the state-space parametric uncertainty and then using experimental results for load step changes.

The experimental comparison also includes the one degree of freedom H loopshaping controller

and the ECU controller provided with the engine as benchmark.

While mixed analysis is part of the Analysis and Synthesis Toolbox, mixed synthesis is not

yet. However, thanks to Peter Young, access to a mixed synthesis algorithm (so-called (D,G)-K

iteration) was made available and used for an additional control design presented in Section 5.6 based

on the parametric state-space uncertainty. Finally, the conclusions of these design studies are drawn

in Section 5.7.

As pointed out in the Section 4.3, there is significant uncertainty in the model of the EGR-VGT

system. A model-based controller must therefore be robust in order to work on the actual engine as

well as in simulation. This section describes an H loopshaping design as described in Appendix A

for the nominal operating point (1500 rpm, 85 Nm, cf. Table 4.1). This methodology comprises two

63

64 Chapter 5. Uncertainty Parameterisations and Robust Control Design

steps: Firstly, the scaling and weighting of the multivariable plant with pre- and postcompensators

to shape the frequency response. Thus, engineers can use their experience and knowledge of the

plant characteristics to obtain a trade-off between performance and robust stability. Secondly, the

stability radius of the shaped plant is optimised with respect to coprime factor uncertainty.

Note that for the EGR-VGT control design, the underlying actuator position control loops from

(4.1) have to be included in the plant model. Figure 5.1 shows the overall setup with all the elements

appearing throughout the subsequent control design.

Shaped Plant Ps

Plant + Act. Pos. Control

K W1 B SC Pact Peng C SC W2

5.1.1 Scaling

Scaling the inputs and outputs of the plant is necessary in order to ensure the appropriate relative

importance of the signals in the cost function to be minimised. As suggested in [SP96], the inputs

are scaled by the largest allowed input change, while the outputs are scaled by the largest allowed

control error, hence normalising the signals to be less than one in magnitude. For the inputs, which

can both range between 0 and 100 % in actuator position, this is achieved by dividing by 100 (note

that the actual range at the nominal operating point is smaller, but the key point is that both channels

are scaled by the same values). For the outputs, the derivations from the setpoints are estimated by

experimentally obtaining the MAF and MAP values at the highest expected engine operating point

(for this control study 1500 rpm, 150 Nm are chosen). This results in a deviation of 25 kg/h in MAF

and of 10 kPa in MAP. Therefore, the input (BSC ) and output (CSC ) scaling matrices are obtained as:

1

100 0 0

B SC = , C SC = 25

1

.

0 100 0 10

Note that these scaling matrices were also applied to the frequency responses in Figures 4.6 and 4.7

in order to make the different channels comparable.

Choice of Weights

The choice of weights requires experience, intuition, and good knowledge of the plant. Since the MAF

and MAP loops have different speeds in response, the magnitude plots are shaped in this application

5.1. H Loopshaping Control 65

rather than the singular values. However, the singular values still determine the robustness properties

of the loop and are subsequently plotted as well. The objective is to have high low-frequency gain

for performance, not too steep slope at cross-over frequency for robustness, and low high-frequency

gain for robustness and sensor noise attenuation. The latter can be achieved by low-pass filters in the

postcompensator W2 , which is typically implemented in the feedback path. W2 could also contain

lead-lag filters to reduce the phase lag at cross-over. However, in this application, the plant rolls off

naturally such that a postcompensator is not necessary. This has the advantage of producing a lower

order controller.

The precompensator W1 is chosen as PI weight for both channels. The integrators are necessary

to achieve zero steady-state error by boosting the low-frequency gain. The zeros of the PI weights are

placed at the intended closed-loop bandwidths (which have been selected as approximately 0.4 Hz

for the EGR loop and 0.2 Hz for the VGT loop) to stop the roll-off and hence to avoid additional

phase lag at cross-over. Finally, constant gains in both channels are chosen such that the cross-

over frequencies of the shaped plant coincide with the targeted closed-loop bandwidths of each

channel. Some design iterations are necessary to achieve the exact closed-loop bandwidths since the

inclusion of K slightly changes the loopshape as discussed later. These considerations yield the

precompensator:

0.015 1s s+20.4 0

W1 = 20.4

. (5.1)

0 0.006 1s s+20.2

20.2

The magnitude plots and loopshapes of the singular values of the scaled and weighted plant are

shown in Figure 5.2. Typically, it is desired to increase the bandwidth of the scaled plant. However,

this is not possible here, since the scaled plant contains the actuator position control loops (1.0 Hz

bandwidth for the MAF channel and 0.5 Hz bandwidth for the MAP channel). Hence, the bandwidth

of the closed-loop system has to be lower than that of the inner actuator loops to avoid undesired

interactions. This is reflected in the targeted bandwidths of 0.4 Hz for the MAF channel and 0.2 Hz

for the MAP channel. In order to render the system response faster, a two degrees of freedom

approach will be taken in Section 5.2.

Controller Synthesis

The controller for the weighted plant is obtained using the Matlab command ncfsyn, which returns

the optimal stability margin from equation (A.3) and the controller achieving it (ncfsyn assumes

positive feedback!). The maximal stability margin for the plant with the aforementioned weights is

max = bopt = 0.65, which is very large (typically, a stability margin of 0.3 is considered sufficient).

This is due to the small targeted bandwidths, which are limited by the actuator position control loops.

The optimal controller K has seven states and therefore one less than the shaped plant, which is

an inherent property of the optimal controller. Note that the plant including the actuator position

control loops has been model-reduced from eighth to sixth order without significant effect on the

frequency response by truncation of the balanced realisation (sysbal); the remaining two states

are introduced by the pre- and postcompensators.

66 Chapter 5. Uncertainty Parameterisations and Robust Control Design

EGR VGT

P

s

Pw

1

0

10

0

10 10

MAF

2 2

10 10 0

10

2 0 2 2 0 2

10 10 10 10 10 10

1

10

0 0

MAP

10 10

2

10

2 2

10 10

2 0 2 2 0 2 2 1 0 1 2

10 10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz f in Hz

Figure 5.2: Magnitude plots (left) and singular values (right) of the scaled (dash) and weighted

(solid) plant.

Controller Analysis

Figure 5.3 shows the magnitude plots and singular values of the synthesised controller. The inclusion

of K has lowered the gain of the singular values at all frequencies. While the main couplings are

hardly affected at all, the gain in the cross-couplings (which is also where the major uncertainty lies

according to Figure 4.7) has changed slightly. Note that bounds on the deterioration of the loopshape

due to the inclusion of the optimal controller can be derived, cf. [ZDG96]. A large robust stability

margin indicates that the effect on the loopshape is small, which is the case here.

EGR VGT

PwK

Pw

1

0

10

0

10 10

MAF

2 2

10 10 0

10

2 0 2 2 0 2

10 10 10 10 10 10

1

10

0 0

MAP

10 10

2

10

2 2

10 10

2 0 2 2 0 2 2 1 0 1 2

10 10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz f in Hz

Figure 5.3: Magnitude plots (left) and singular values (right) of the weighted plant (solid) and with

K in the loop (dashed).

5.1. H Loopshaping Control 67

Controller Implementation

The controller can be implemented in the forward path (small rise time from reference to output, but

overshoots as high frequency controller dynamics are excited by the references), the feedback path

(slower, more damped response), or in the observer form, which is the optimal way of introducing the

controller to meet the performance specifications [PG97]. In this study, the controller is implemented

in the feedback path with the weight W1 in the forward path. In order to avoid windup of the integrator

states in W1 due to actuator constraints, it is implemented in its self-conditioned or Hanus form. This

implementation keeps the states of W1 consistent with the actual plant input at all times; cf. [HKH87]

for more details on this anti-windup scheme.

Since the inclusion of K in the loop has lowered the system gain, it needs to be checked whether

the targeted closed-loop bandwidths are achieved. Figure 5.4 shows the frequency responses of

the closed-loop transfer functions with the weights moved from the plant into the controller. The

targeted 3 db bandwidths of the EGR loop (0.4 Hz) and of the VGT loop (0.2 Hz) are met almost

exactly (admittedly, this has only been achieved after some design iterations). The final step is now

to validate the controller in (nonlinear) simulation and experiment.

10

0 0

10 10

MAF

0

10

2

10

2

10

2 0 2 2 0 2

10 10 10 10 10 10

1

10

0 0

10 10

2

10

MAP

2 2

10 10

3

2 0 2 2 0 2 10

10 10 10 10 10 10 2

10

1

10 10

0 1

10

2

10

f in Hz f in Hz f in Hz

Figure 5.4: Magnitude plots (left) and singular values (right) of the closed loop with weights moved

into the controller.

For the controller validation, steps in the MAF and MAP reference around the nominal operating

point values are chosen (84 kg/h, 111 kPa, cf. Table 4.1). Figure 5.5 shows the reference, experimen-

tal, and nonlinear simulation data. The plots on the left side depict the response to a step in MAF from

74 to 94 kg/h and down again, while MAP is kept constant. The simulated and experimental data

coincide almost perfectly for the MAF channel, but the experimental response in the MAP channel

is significantly underestimated. However, this is not surprising since the gain of this cross-coupling

is underestimated in the model as shown in Figure 4.6.

68 Chapter 5. Uncertainty Parameterisations and Robust Control Design

100 100

ref

95 exp 95

sim

90 90

MAF in kg/h

85 85

80 80

75 75

70 70

0 5 10 15 20 25 30 0 5 10 15 20 25 30

120 120

115 115

MAP in kPa

110 110

105 105

0 5 10 15 20 25 30 0 5 10 15 20 25 30

t in sec

t in sec

Figure 5.5: Validation of the H loopshaping controller with steps in the MAF and MAP references

(experiment and nonlinear simulation).

The plots on the right side of Figure 5.5 show the response of the control system to a step in

the MAP reference, while MAF is kept constant. From the scaling in Section 5.1.1, it has been

established that a change of 1 kPa in MAP is equally significant as a change in MAF of 2.5 kg/h.

Hence, the step height for the MAP reference has been chosen as 8 kPa. Again, the agreement in the

main coupling (bottom right plot) is very good, while the gain of the cross-coupling (top right plot)

is again underestimated but not as severely as for the EGR to MAP channel (bottom left plot).

Note that for the controller validation in Figure 5.5, only the reference values of MAF and MAP

are fed into the simulation model. The agreement of the simulation and experimental data gives

confidence that a controller tuned on the model should give a similar performance in experiments.

Moreover, note that the experimental data have been low-pass filtered with a cut-off frequency of

1.5 Hz for the presentation of the results.

The achievable bandwidth with the one degree of freedom loopshaping controller is limited by the

underlying actuator position control loops. If the outer loop becomes to fast, the interaction between

the controllers might cause instabilities. However, the closed-loop bandwidth can be increased by

adding a controller in the feedforward path, which cannot cause instabilities. In such a controller, the

commands and feedbacks are processed independently (resulting in two degrees of freedom) rather

than having an error signal as input.

5.2. Two Degrees of Freedom Controller for Coprime Uncertainty 69

A two degrees of freedom extension of the H loopshaping methodology was proposed by Limebeer

et al. [LKP93]. It allows for the design of a feedback controller to robustly stabilise the plant with

respect to coprime factor uncertainty and of a dynamic prefilter to make the closed loop match a

reference model in a single step. The design configuration is depicted in Figure 5.6, which is an

N ) are assumed to be

extension of the standard framework in Figure A.7. The transfer matrices (M,

a left coprime factorisation of the shaped plant, i. e. Ps = M 1 N . Note that is a tuning factor that

weighs the relative importance of the model matching against the robust stabilisation.

z1 z2

N

M

w

r I u y

I e

K1 N M 1

K2

Tre f

In order to apply the standard Matlab algorithms for the robust performance problem at hand, it

needs to be cast into the general framework as described in Appendix A.1.6. With the exogenous

inputs being (w, r) and the performance signals being (z1 , z 2 , e), the generalised plant P relating

these signals can be read off the block diagram in Figure 5.6:

z1 0 0 I

1

z2 M 0 Ps w

e = M 1 2 Tre f Ps

r

0 I 0 u

y M 1

0 Ps

with the zero and unity submatrices of appropriate dimensions. This robust performance problem

can now be reformulated as an equivalent robust stability problem with augmented uncertainty as

depicted in Figure 5.7, where

u

P :=

0

0

f

: u 24 , f 22

contains two blocks, u for the robust stabilisation as in the standard H loopshaping methodology

and f for the model matching with the reference model Tre f .

70 Chapter 5. Uncertainty Parameterisations and Robust Control Design

u

u

f

z1, z2 w z1, z2, e w, r

e r

P P

, y u , y u

K K

Since the two degrees of freedom design is an extension of the loopshaping procedure from Section

5.1, the same scaling and precompensator are chosen. It now remains to specify a reference model

for the model matching of the closed loop. A reasonable requirement is to make the overall system

as fast as the actuators. The bandwidths of the actuator position control loops were identified in (4.1)

as 1.0 Hz for the EGR loop and 0.5 Hz for the VGT loop. Hence, the reference model is chosen as

21.0

0

Tre f = s+21.0

20.5

. (5.2)

0 s+20.5

The cross-couplings in the reference model are chosen to be zero to decouple the channels; the

steady-state gain is unity as required by reference tracking.

Controller Synthesis

Due to the structure in the uncertainty p , the optimisation becomes the following synthesis

problem:

K stabilising

which can be tackled by the so-called D-K iteration technique (Matlab command dkit) as mentioned

in Appendix A.3.1. This procedure involves two minimisations, one over scaling matrices D (which

are fitted as dynamic systems of prespecified order) to obtain and one over the set of stabilising

controllers K , which is done iteratively by solving for K and D separately while keeping the other

variable constant. Although global convergence is not guaranteed, many designs have shown that

this approach often works well in practice. Note that for the problem at hand, with a structured

uncertainty consisting of two full blocks, not only the minimisation over K is convex but also over

D. Nevertheless, the joint optimisation over D and K is not convex.

5.2. Two Degrees of Freedom Controller for Coprime Uncertainty 71

After some iterations and with a tuning factor of = 1, a value of 1.56 is achieved. The

synthesised controller K = [K1 , K 2 ] has ten states and therefore as many as the generalised plant,

which has the same eight states as in the previous loopshaping design plus two states from the

reference model. Note that K1 and K 2 share the same state space. For this and the subsequent

synthesis designs, the scaling matrices D are chosen to be static to keep the controller order as low

as possible (the scaling matrices are absorbed into the generalised plant during the synthesis leading

to increased controller order when increasing the order of their dynamic fit). Note that increasing

the order of the scalings did not significantly improve the achieved values.

Controller Analysis

In order to assess the robustness with respect to coprime factor uncertainty, the H norm of the

closed-loop transfer function from w to z1 , z 2 is determined. It turns out to be identical to the

value of 1.56. The inverse of this (1/1.56 = 0.64) can now be compared to the robust stability margin

obtained with the H loopshaping controller from Section 5.1, which was found to be bopt = 0.65.

Hence, the two degrees of freedom controller achieves almost the same robustness. However, this is

not surprising, since the additional controller K1 is not in the feedback loop and thus does not affect

the stability. With this result, a fair comparison of the performance of these two controllers is now

possible.

Controller Implementation

Figure 5.8 depicts the implementation setup for the two degrees of freedom controller. As mentioned

before, K1 and K 2 share the same state space, and should hence be implemented in one block. Wi

is a constant scaling matrix required to make the feedback steady-state gain match the gain in the

forward path, i. e.

Wi = K 11 (0)K 2 (0)C SC .

The precompensator W1 , which includes the integrators, is again implemented in the Hanus anti-

windup scheme.

r y

Wi [K 1 , K 2 ] W1 B SC Pact Peng

C SC

Figure 5.8: Implementation setup for the two degrees of freedom controller.

72 Chapter 5. Uncertainty Parameterisations and Robust Control Design

Figure 5.9 shows the frequency responses of the closed-loop transfer functions for the two degrees

of freedom coprime uncertainty controller (CUC, solid lines). The comparison to the closed-loop

frequency responses with the H loopshaping controller, which are here plotted again (dashed lines),

shows that the bandwidths of the system have indeed been increased, especially in the VGT to MAP

channel. However, compared to the frequency response of the reference model, the 2DOF controller

rolls off faster at higher frequencies. The model matching can be improved by tuning the value, but

this results in reduced robustness such that a fair performance comparison to the H loopshaping

controller would not be possible anymore.

MAF Ref MAP Ref 1

10

CUC

0 0 H

10 10

Tref

MAF

0

10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10 1

10

0 0

10 10

2

Tref 10

MAP

2 2

10 10

3

10 2 1 0 1 2

2 0 2 2 0 2

10 10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz f in Hz

Figure 5.9: Magnitude plots (left) and singular values (right) of the closed loop with the two degrees

of freedom coprime uncertainty (solid) and the H loopshaping controller (dashed). The left plots

also show the frequency response of the reference model (dotted).

For the experimental controller validation, the same MAF and MAP reference steps are chosen

as in Figure 5.5. The simulation and experimental results are depicted in Figure 5.10, which also

includes the previous results for the H loopshaping controller (dotted lines). As expected, the re-

sponse time of both channels has become significantly smaller resulting in an improved performance.

The experimental and simulation data coincide very well for the MAF channel. With respect to the

MAP channel, the response to a step in the MAF reference is almost identical to the H controller

response. However, the response to the step in MAP (bottom right plot) shows some overshoot in

the experiment which does not occur in the simulation. This is most likely again due to the underes-

timation of the gain in the cross-coupling from EGR to MAP, which now becomes apparent due to

the increased bandwidth. The forward part of the two degrees of freedom controller, i. e. K1 , relies

on a good model. Due to the gain underestimation, it leads to an overshoot, which is then corrected

by the slow feedback part of the controller, i. e. K2 .

Finally, the response in the cross-coupling from the MAP reference to MAF has become larger,

in simulation as well as in experiment. This can be explained by the increased closed-loop gain in

that cross-coupling as shown in Figure 5.9 (left side, upper right plot).

5.3. Two Degrees of Freedom Controller for Additive Uncertainty 73

100 100

ref

95 sim 95

exp

H

MAF in kg/h

90 90

85 85

80 80

75 75

70 70

0 5 10 15 20 25 30 0 5 10 15 20 25 30

120 120

115 115

MAP in kPa

110 110

105 105

0 5 10 15 20 25 30 0 5 10 15 20 25 30

t in sec t in sec

Figure 5.10: Validation of the two degrees of freedom coprime uncertainty controller with steps in

the MAF and MAP references (experiment and nonlinear simulation). The dotted lines correspond

to the experimental validation data of the H loopshaping controller, cf. Figure 5.5.

As described in Section A.2.2, the H loopshaping methodology optimises the robustness of the

system with respect to coprime factor uncertainty. The reason for using this class of uncertainty is

that it is very general and includes a wide class of other unstructured uncertainty types, e. g. additive

and multiplicative uncertainty. The question to be addressed in this section is how to incorporate an

uncertainty description tailored to the application, i. e. taking advantage of the extra knowledge from

Figure 4.7, which shows where the main uncertainties occur in the model of the EGR-VGT system.

based on the frequency response and error bounds in Figure 4.7 directly. In order to make such a

control design comparable to the H loopshaping controller from Section 5.1 and the two degrees

of freedom coprime uncertainty controller from Section 5.2, the same weights are to be used. The

weighted plant is then augmented with an additive uncertainty as shown in Figure 5.11, which

replaces the coprime factor uncertainty in Figure 5.6. The signals w2 and w3 are required to arrive

at a sensible optimisation problem. Note that due to the equivalence of (A.1) and (A.2), the signal

w in Figure 5.6 fulfils the same function as the signals w2 and w3 in Figure 5.11.

74 Chapter 5. Uncertainty Parameterisations and Robust Control Design

z1 w1

W1 u W2 I

w2 w3

r u y e

I K1 Ps I

K2

Tre f

The plant plus uncertainty in Figure 5.11 represent a class of models P obtained by adding

weighted uncertainty to the nominal shaped model Ps :

u (s)

< 1,

where W1 and W2 are stable transfer matrices that characterise the spatial and frequency structure

of the uncertainty ( I in Figure 5.11 is a tuning factor for the design, see below, and therefore

omitted in the equation). Note that due to the additive nature, the same uncertainty description of the

unshaped plant would be different, i. e. the loopshaping weights affect the uncertainty description,

which is not true in the case of multiplicative uncertainty. However, in this application with the

uncertainties mainly occurring in the cross-couplings, the additive uncertainty description is easier

to obtain (since matrix multiplication is not element by element).

Figure 5.12 shows the frequency responses of the nominal shaped plant (black lines) and of

the shaped plant with varied parameters (red lines). Again, the uncertainty is largest in the low

frequency range of the cross-couplings. In order to construct the uncertainty weights W1 and W2 ,

the uncertainty bounds are obtained for the system by plotting the error

err ( j ) = P ( j ) Ps ( j ),

where Ps corresponds to the nominal (shaped) frequency response (black traces in Figure 5.12)

and P to the frequency responses with worst-case parameters (red traces, upper and lower limit)

implying

u (s)

= 1. This approach assumes that the frequency responses for smaller parameter

variations are captured by

u (s)

< 1, which seems reasonable. The additive error err then has

to be bounded by fitted transfer functions, subsequently referred to as matrix Wf it . A good fit can

be obtained by second and third order transfer functions. Using these fits, the uncertainty bounds on

the shaped plant are plotted in Figure 5.12 (blue lines). The good agreement gives confidence that

they describe the uncertainty rather well.

5.3. Two Degrees of Freedom Controller for Additive Uncertainty 75

2 EGR 2 VGT

10 10

0 0

10 10

MAF 2 2

10 10

model

uncert

bounds

4 4

10 10

2 0 2 2 0 2

10 10 10 10 10 10

2 2

10 10

0 0

10 10

MAP

2 2

10 10

4 4

10 10

2 0 2 2 0 2

10 10 10 10 10 10

f in Hz f in Hz

Figure 5.12: Frequency response of the shaped plant with varied parameters and additive uncertainty

bounds.

For the control design, the fitted additive uncertainty has to be factorised as W2 u W1 :

1 1 0

1 Wfit 2 Wfit Wfit Wfit 0 0 2 0 1

11 12 = 11 12

3 Wfit 4 Wfit 1 0

21 22

0 0 Wfit

21

Wfit

22

3

4 0 1

:= W2 (s)u (s)W1 (s),

u (s)

< 1, i .

The four different scalars 1 , . . . , 4 imply that each element in W f it changes independently as u

varies. Since the uncertainty that originally arose from three real parameters is now covered by four

complex parameters, it implies that this parameterisation is rather conservative. Nevertheless, it will

turn out to be useful for control design.

As mentioned before, the weight W1 and the reference model Tre f are both chosen identical to the

two degrees of freedom coprime uncertainty design in Section 5.2. The tuning factor , which

again weighs the relative importance of the model matching against the robust stabilisation, will be

adjusted iteratively during the design. The additional tuning factor in Figure 5.11 is introduced

76 Chapter 5. Uncertainty Parameterisations and Robust Control Design

to directly adjust the amount of additive uncertainty considered in the design. This will be further

discussed subsequently.

Controller Synthesis

As for the previous two degrees of freedom design, the control problem at hand has to be cast into

the standard framework. Firstly, the generalised plant P is read off the block diagram in Figure 5.11,

with the exogenous inputs being (w1 , w2 , w3 , r) and the controlled signals being (z1 , u, e) in this

case:

z1 0 W1 0 0 W1 w1

u 0 I 0 0 I w2

e = W Ps I 2 Tre f Ps w3

2

0 0 0 I 0 r

y W2 Ps I 0 Ps u

with the zero and unity submatrices of appropriate dimensions. Again, this robust performance

problem can now be reformulated as an equivalent robust stabilisation problem with augmented

uncertainty (cf. Figure 5.7, but with the new signals of course), where now

u

P :=

0

0

f

: u = diag(1 , 2 , 3 , 4 ), i , f 64 .

With this setup, the D-K iteration is applied with the tuning factors = = 1. However, it turns out

that the resulting controller is too conservative, i. e. the performance is very poor in order to guarantee

stability for the parameterised uncertainty. To achieve a control design which is comparable to the

designs from the previous sections, the tuning factors are adjusted until the performance in the

nonlinear simulation of the two degrees of freedom controller from Section 5.2 is met. This is

achieved with = 2 and = 0.5. Figure 5.13 shows the time responses to steps in MAF and MAP

in nonlinear simulation for the two controllers. While the MAF traces are indeed almost identical,

there are slight differences in the MAP response. However, both controllers achieve the same 0 to

63 % rise time and reach the final value simultaneously.

The value achieved with the aforementioned choice of tuning factors is 3.73. The synthesised

controller K = [K1 , K 2 ] has 16 states, again as many as the generalised plant, which has been

augmented by six states due to the uncertainty bounds. Again, K1 and K 2 share the same state space.

Controller Analysis

To assess the achieved robust stability, the H norm of the closed-loop transfer function from w1 to

z 1 is calculated to 0.42. Considering the downtuning with = 0.5, this implies a gain of 0.84 from

w1 to z 1 and indicates that robust stability is achieved. The analysis of the system in Section 5.5.1

will give further insight.

5.3. Two Degrees of Freedom Controller for Additive Uncertainty 77

95

AUC

CUC

90

Ref

MAF in kg/h

85

80

75

70

0 5 10 15 20 25 30

116

114

MAP in kPa

112

110

108

106

0 5 10 15 20 25 30

t in sec

Figure 5.13: Performance comparison of the additive uncertainty (AUC) and the coprime uncertainty

(CUC) controller.

Controller Implementation

The controller designed for additive uncertainty can be implemented in exactly the same way as the

two degrees of freedom controller in the previous section (cf. Figure 5.8).

Figure 5.14 shows the closed-loop frequency responses with the additive uncertainty controller (AUC)

in comparison to the coprime uncertainty controller (CUC) from Section 5.2. The main couplings

are very similar for both controllers and roll off slightly faster than the reference model. However,

the gains in the cross-couplings (especially in the VGT to MAF channel) are smaller for the additive

uncertainty design resulting in better decoupling.

10

CUC

0 0 AUC

10 10

Tref

MAF

0

10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10

1

10

0 0

10 10

2

Tref 10

MAP

2 2

10 10

3

2 0 2 2 0 2 10

10 10 10 10 10 10 10

2 1

10 10

0 1

10

2

10

f in Hz f in Hz f in Hz

Figure 5.14: Magnitude plots (left) and singular values (right) of the closed loop with the addi-

tive uncertainty (solid) and coprime uncertainty controller (dashed). The left plots also show the

frequency response of the reference model (dotted).

78 Chapter 5. Uncertainty Parameterisations and Robust Control Design

Figure 5.15 shows the time domain validation data with respect to the same step excitations

as before. While there is hardly any change compared to the coprime uncertainty controller from

Section 5.2 (dotted magenta lines) in the MAF channel, the overshoot in the MAP channel could be

reduced significantly without sacrificing the performance. This is due to the decreased closed-loop

gain in the cross-couplings as shown in Figure 5.14. Note that the measured data have again been

low-pass filtered with a 1.5 Hz cut-off frequency for the presentation of the results.

100 100

ref

95 sim 95

exp

CUC

MAF in kg/h

90 90

85 85

80 80

75 75

70 70

0 5 10 15 20 25 30 0 5 10 15 20 25 30

120 120

115 115

MAP in kPa

110 110

105 105

0 5 10 15 20 25 30 0 5 10 15 20 25 30

t in sec t in sec

Figure 5.15: Validation of the additive uncertainty controller with steps in the MAF and MAP

references (experiment and nonlinear simulation). The dotted lines correspond to the experimental

validation data of the coprime uncertainty controller, cf. Figure 5.10.

This experimental validation indicates that the parameterisation of the identified uncertainty

does improve the controller performance. It should be noted, however, that the additive uncertainty

controller is of order 16 as compared to order ten of the two degrees of freedom controller. Further

validation results are provided in Section 5.5.

Two alternative uncertainty parameterisations are discussed in this section, both leading to descrip-

tions in which the parameters are real. While the first approach, state-space parametric uncertainty,

results in a valid description of the uncertainty at hand, the second does not. Since the controller

synthesis for real parametric uncertainty is not yet included in the Matlab Analysis and Synthesis

5.4. Alternative Uncertainty Parameterisations 79

Toolbox, the parameterisation derived in the next section will initially be used for the analysis of

the previously designed controllers in Section 5.5.1.

Ideally, the uncertainty would be parameterised directly based on the parameters in the nonlinear

model that are known to be most uncertain, i. e. the effective area of the EGR valve and the efficiencies

of the turbine and compressor. However, these parameters (especially the turbocharger efficiencies)

are rather complicated polynomial functions of the inputs and states of the system. Hence, the

linearised model is used subsequently to obtain an uncertainty description.

For this approach, the nonlinear model is linearised for separate changes in each of the three param-

eters. Their effect on the nominal state-space model (A, B, C, D) is then expressed as follows

3

3

A + i=1 i A i B+

i Bi

G (s) =

i=1 ,

3 3

C+ i C i D+

i Di

i=1 i=1

where i [1, 1] and the matrices Ai , B i , C i , D i reflect how the ith uncertainty, i , affects the state-

space model. This uncertainty description can then be reformulated to obtain an LFT representation

as described e. g. in [ZDG96]. In this case, the uncertainty block is obtained as

u = diag{1 I4 , 2 I2 , 3 I1 },

where 1 corresponds to the effective EGR valve area, 2 and 3 to the compressor and turbine

efficiency, respectively. Thus, the uncertainty is expressed by three real parameters rather than by

four complex ones as in the case of the additive uncertainty description.

However, it remains to confirm that a combination of the uncertainty parameters results in a valid

uncertainty description. This is done by plotting the frequency response for simultaneously changed

parameters. A comparison of Figure 5.16 to Figure 4.7 proves the validity of this approach.

As mentioned before, mixed complex and real synthesis is not yet part of the Matlab Analysis

and Synthesis Toolbox. Therefore, the parametric state-space uncertainty description is only used

for the analysis of the previously designed controllers in Section 5.5.1. The extension to mixed

synthesis will then be discussed in Section 5.6.

An alternative solution is to use the fact that the turbocharger efficiencies mainly affect the tur-

bocharger speed, which is a state of the model, and the effective area is a static nonlinear function

of the VGT actuator position and can thus be regarded as a model input. Then, using the linearised

model, it is possible to extract the uncertainties of the turbocharger speed state and the effective area

input; a process, which is sometimes referred to as pulling out the deltas [ZDG96]. From nonlinear

simulations it can be derived how much the turbocharger speed varies with changes in the efficien-

80 Chapter 5. Uncertainty Parameterisations and Robust Control Design

1

EGR [%] 1

VGT [%]

10 10

0 0

10 10

MAF

1 1

10 10

nominal

limits

0 2 0 2

10 10 10 10

1 1

10 10

0 0

10 10

MAP

1 1

10 10

2

10

2 0 2 0 2

10 10 10 10 10

f in Hz f in Hz

Figure 5.16: Frequency response of the scaled plant at 1500 rpm and 85 Nm and the limits with

worst-case parametric state-space uncertainty.

cies. The same can be done for the effective area and the actuator position. Thus, a linearised model

with parametric uncertainty is constructed.

However, comparing the frequency responses obtained by assuming maximal deviation of the

parameters in Figure 5.17 to the results for maximal parameter changes in the nonlinear model

(cf. Figure 4.7) it turns out that the linear uncertainty description does not capture the effects in the

nonlinear model, i. e. the nonlinear changes are too significant to be captured by this linear approach.

Note that the EGR input in Figure 5.17 is the effective area of the valve, not the actuator position,

hence the scaling had to be adjusted.

In this section, the two degrees of freedom controllers (CUC and AUC) are analysed based on a real

analysis with respect to the state-space uncertainty description. Then, all controllers designed in

this chapter are compared to the ECU controller provided with the engine for a realistic load step

event.

5.5. Controller Comparison 81

2

EGR A [m ] VGT [%]

r

1 1

10 10

0 0

MAF 10 10

1 1

10 10

limits

nominal

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10

1 1

10 10

0 0

10 10

MAP

1 1

10 10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10

f in Hz

f in Hz

Figure 5.17: Frequency response of the scaled plant at 1500 rpm and 85 Nm and the limits with

worst-case pulled out deltas. Note that the EGR input is the effective area of the valve, not the

actuator position.

The CUC and AUC are now analysed based on the parametric state-space uncertainty description

from Section 5.4.1 containing three real parameters and therefore being the least conservative. The

setup for the analysis is shown in Figure 5.18. The robustness and performance are analysed via the

transfer functions from w to z and r to e, respectively. Additional signals, which were necessary for

a sensible controller synthesis setup, can now be omitted. Note that Wi is again a constant scaling

matrix to make the closed-loop transfer function from r to y match the reference model exactly at

steady-state, i. e. Wi = K 11 (0)K 2 (0) in this case where the output scaling matrix CSC is absorbed

in the plant.

The results of the real analysis are shown in Figure 5.19. The plot (left) shows that the

peak value is reduced by 46 % from 1.16 for the CUC to 0.55 for the AUC. The plot on the

right side of Figure 5.19 shows the graphs for robust stability (i. e. the maximal singular value of

the transfer function from w to z) and nominal performance (i. e. the maximal singular value of the

transfer function from r to e). It reveals that the large value for the CUC is due to the nominal

performance, while the robust stability is almost exactly the same for both controllers. This indicates

82 Chapter 5. Uncertainty Parameterisations and Robust Control Design

u

z w

r Ps e

Wi K1

K2

Tref

that for the tighter additive uncertainty description, the minimisation can focus on the performance.

Note that the robust stability for the AUC is 0.46 for the state-space uncertainty as compared to 0.84

for the additive uncertainty (cf. Section 5.3.2). Hence, the state-space uncertainty description is less

conservative as expected.

Plot

Robust Stability and Nominal Performance

1.2

1.2

AUC

AUC

CUC

CUC

1

1

0.8

0.8

0.6

0.6

Robust Stability Nominal Performance

0.4

0.4

0.2

0.2

0 2 1 0 1 2 0 2

10 10 10 10 10 1 0 1 2

10 10 10 10 10

f in Hz f in Hz

Figure 5.19: Comparison of plots (left) and robust stability as well as nominal performance (right)

of the additive uncertainty (solid) and coprime uncertainty controller (dashed).

The large gain in the performance block could also indicate that the reference model was badly

chosen. However, comparing Figure 5.19 to Figure 5.14 shows that the peaks in the nominal per-

formance (and also in the plot) at around 1 Hz are due to the cross-couplings, which are meant

to match the zero gain of the cross-couplings in the reference model. Thus, a lot of control effort

goes into this task, but that is sensible because the decoupling is desired. In conclusion, the tighter

additive uncertainty description allows better decoupling.

5.5. Controller Comparison 83

The performance of the three controllers previously designed (H , CUC, AUC) will now be com-

pared to the gain scheduled PI controller with feedforward implemented on the engine. Figure 5.20

depicts the responses with respect to a load change of 35 Nm around the nominal operating point.

The reference values for MAP and MAF are stored in an ECU map based on speed and fuelling and

hence change simultaneously.

110 125

100 120

MAF in kg/h

MAP in kPa

90 115

80 110

70 105

60 100

0 10 20 30 40 50 0 10 20 30 40 50

Zoom Zoom

125

100

120

MAF in kg/h

90

MAP in kPa

115

ECU

80 Ref

110

2DOF

H

70 AUC 105

60 100

29 30 31 32 33 34 29 30 31 32 33 34

t in sec t in sec

Figure 5.20: Comparison of ECU to H , CUC, and AUC controller with respect to load changes

(85 35 Nm).

The different responses in Figure 5.20 confirm the observations from the previous validations.

The H controller is very robust but rather slow, especially in the MAP response (green lines).

The coprime uncertainty controller with two degrees of freedom increases the bandwidths of both

channels, but causes significant overshoot in the MAP channel and oscillatory behaviour in the MAF

channel (blue lines). The same response times but with less overshoot and oscillation is achieved

with the additive uncertainty controller (magenta lines), which reduces the mean-square tracking

error by 22 %. This again confirms the suitability of this approach. The provided ECU controller

shows a very slow response in the MAP channel, and does not track the MAF reference very well at

the high load point (red lines). However, the MAF response is as fast as with the other controllers

and almost without oscillation.

Note that the gain in the VGT channel changes significantly when increasing the load (i. e. in-

creased exhaust gas energy results in more control authority of the VGT). The calibrated ECU

84 Chapter 5. Uncertainty Parameterisations and Robust Control Design

controller compensates for this by a change in the I gain of the controller, while the other controllers

purely rely on their robustness.

Thanks to Peter Young, an extension to the existing synthesis (which only covers complex pa-

rameters) to mixed real and complex problems was made available. The algorithm is based on the

so-called (D,G)-K iteration, which includes a G scaling to exploit the phase information given by the

real parameters. Details can be found in [You01]. As for the D-K iteration, there are no convergence

guarantees.

The (D,G)-K iteration will be used in this Section to design a controller based on the parametric

state-space uncertainty, for which the CUC and AUC were already analysed in Section 5.5.1. The

state-space parameterisation describes the uncertainty by three real parameters, as opposed to four

complex ones for the additive uncertainty. A less conservative control design is therefore expected.

The setup for the design is very similar to the setup for the additive uncertainty controller depicted

in Figure 5.11; the signals are the same, only the uncertainty description has changed from additive to

LFT form. Again, choosing the tuning factors as = 2 and = 0.5 achieves the same performance

in time domain simulations as for the CUC and AUC in Figure 5.13. Figure 5.21 compares the

controller synthesised for state-space uncertainty (SSUC) to the CUC and AUC. Due to the less

conservative uncertainty description, the performance could again be increased.

1.2

1.2

CUC

AUC

AUC

CUC

SSUC

SSUC 1

1

0.8

0.8

0.6

0.6

Robust Stability Nominal Performance

0.4

0.4

0.2

0.2

0

0 2 1 0 1 2

1 0 1 2 10 10 10 10

10 10 10 10 10

f in Hz f in Hz

Figure 5.21: Comparison of plots (left) and robust stability as well as nominal performance (right)

of the state-space uncertainty (magenta) to the additive uncertainty (solid) and coprime uncertainty

controller (dashed).

The main improvement in the plots for the SSUC is obtained by having one parameter less.

The further restriction to those parameters being real rather than complex only results in very little

further improvement. This is due to the fact that the peak value occurs at around 1 Hz where the

robust stability is already excellent and the peak is due to the nominal performance block, which

corresponds to the complex uncertainty block f .

5.7. Conclusions 85

1

10

0 0

CUC

10 10

AUC

Tref

SSUC

MAF

0

10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10 1

10

0 0

10 10

2

Tref 10

MAP

2 2

10 10

3

10 2 1 0 1 2

2 0 2 2 0 2

10 10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz f in Hz

Figure 5.22: Magnitude plots (left) and singular values (right) of the closed loop with the state-space

uncertainty (magenta), additive uncertainty (solid) and coprime uncertainty controller (dashed). The

left plots also show the frequency response of the reference model (dotted).

Figure 5.22 shows the closed-loop frequency response for the SSUC in comparison to the closed-

loops with the CUC and AUC. As expected from the analysis in Figure 5.21, the magnitude plots

for the AUC and SSUC are very similar. This is also reflected in the simulation and experimental

results provided in Figure 5.23, which shows the response to steps in MAF and MAP. A comparison

of the traces for the SSUC experimental data (red) to the AUC (magenta) data reveals hardly any

difference at all.

5.7 Conclusions

This chapter has shown that model-based controllers are well capable of controlling the EGR-VGT

system at the chosen operating point. In order to cope with the uncertainty in the model identified

in Chapter 4, a standard H loopshaping controller has been designed. Its stability margin is very

large due to the bandwidth limitations of the actuators. Comparing the simulation and experimental

data shows good agreement, hence, the model-based control design has been successful. In order

to decrease the response time of the control system, an extended loopshaping approach resulting

in a two degrees of freedom controller has been taken. The bandwidth could indeed be increased

achieving the same robust stability margin, but due to the feedforward part relying on a good model,

the underestimated gains in the cross-couplings lead to overshoot especially in the MAP channel.

This could be reduced by designing controllers based on tailored uncertainty descriptions (additive

and state-space parametric). The resulting controllers achieve a similar fast response but with less

overshoot due to decreased gain in the closed-loop cross-couplings.

A further validation step showed that the two degrees of freedom controllers achieve a signifi-

cantly faster response time in the MAP channel than the provided ECU controller, but at the expense

of some overshoot. However, this should only be a problem at very high loads, where the pressure

86 Chapter 5. Uncertainty Parameterisations and Robust Control Design

100 100

exp

95 sim 95

ref

MAF in kg/h

90 AUC 90

85 85

80 80

75 75

70 70

0 5 10 15 20 25 30 0 5 10 15 20 25 30

120 120

115 115

MAP in kPa

110 110

105 105

0 5 10 15 20 25 30 0 5 10 15 20 25 30

t in sec t in sec

Figure 5.23: Validation of the state-space uncertainty controller with steps in the MAF and MAP

references (experiment and nonlinear simulation). The dotted lines correspond to the experimental

validation data of the additive uncertainty controller.

might exceed the prescribed engine limits. The performance in the MAF channel is similar. While

the ECU controller is of significantly lower order, it should be pointed out that it has been calibrated

on the engine (not only the P and I gains, but also the feedforward plus additional operating point

dependent limitations on the I term). Hence, although the structure of the ECU controller is rather

simple, the calibration task is significant. On the other hand, the ECU controller already covers the

whole engine operating envelope. In order to achieve this with the model-based controllers, several

of them have to be designed at different speed-load points. They can then be merged using a gain

scheduling approach. However, there are no guarantees for the robustness of the overall gain sched-

uled controller. Hence, a different approach will be taken in the remainder of this thesis, namely

linear parameter-varying (LPV) gain scheduling. This requires an LPV model, which will be derived

in the next chapter.

From an industrial point of view, controller robustness is a big issue, since a complete recalibration

of the control system on a similar type of engine should be avoided. It must also be ensured, that the

controller calibrated on a particular type of engine can cope with production variability. The design

study in this chapter has shown that H loopshaping and synthesis provide powerful tools to

achieve robust controllers. Moreover, these design methodologies are model-based and hence allow

for a structured approach to controlling the plant and therefore a potential reduction in calibration

effort (avoiding the calibration of different maps without really knowing the cross-effects). The

5.7. Conclusions 87

short-coming of these model-based controllers is their lack of calibratability for fine-tuning on the

engine, which is desired by calibrators during test drives in the vehicle.

Finally, a model-based controller can only be as good as the model. For such complex systems

as turbocharged diesel engines, low-order mean-value models can still be improved as shown by the

gain underestimation in the cross-couplings of the model that led to overshoots in the experimental

MAP response. Another motivation for improving mean-value models is that they can also be used

for controller tuning of simple PI controllers, avoiding the computational burden of complex CFD

models. Thus, the calibration engineer can pretune the controller on his laptop.

88 Chapter 5. Uncertainty Parameterisations and Robust Control Design

6

Linear Parameter-Varying Modelling for Diesel Engines

The controllers in the previous chapter were all designed for a fixed operating point, namely 1500 rpm

and 85 Nm. Due to their robustness, they work properly in a certain region around this operating point,

but their performance will degrade eventually because the system dynamics change significantly with

speed and load, cf. [KMN+97] and [MNK97]. As pointed out in Section 5.7, several controllers

could be designed based on linearised models at different operating points. They would then have

to be merged based on a gain scheduling approach for which there are no robustness guarantees

anymore, although it often works well in practice. The interested reader is referred to [SA91a] for a

discussion of the potential hazards of gain scheduling.

In contrast to the classical gain scheduling techniques, recently developed robust gain scheduling

methods provide systematic ways to design controllers that are scheduled on the operating point of

the system [AA98]. One such approach will be investigated in the next chapter. Since it requires

a linear parameter-varying (LPV) model of the plant, the derivation of such a model for the airpath

control of diesel engines will be discussed subsequently.

LPV systems, which where initially studied in [SA91b], are linear systems whose describing

matrices depend on an exogenous time-varying parameter vector (t):

y = C((t))x + D((t))u

where A, B, C, D are continuous matrix valued functions of (t), which varies in the set of contin-

uously differentiable parameter curves : [0, ) k . Both (t) and its rate of variation are

contained in prespecified compact sets and d . The parameter vector is composed of different

real parameters i each one of them varying between i and i :

i (t) i , i , t 0, i = 1, . . . , k. (6.2)

i (t) q i , q i , t 0, i = 1, . . . , k. (6.3)

The exogenous parameter (t) is unknown apriori but can be measured or estimated online. This

distinguishes LPV systems from linear time-varying systems for which the time-variations are known

89

90 Chapter 6. Linear Parameter-Varying Modelling for Diesel Engines

beforehand. If the scheduling parameter is endogenous to the state dynamics, e. g. (t) is a state

itself as it will be the case for the airpath model, the system shall be called quasi-LPV [SC93].

Note that the parameters are allowed to enter the system matrices in a nonlinear way. However,

it is typically rather difficult to convert a complex nonlinear model into (quasi-) LPV form, although

many nonlinear systems of interest can be written in that form. One of the problems with the engine

model derived in Chapter 3 is that the inputs (EGR and VGT position) enter the A matrix in a

nonlinear way, e. g. via the turbine efficiency maps. Moreover, the order of the system should be as

small as possible because the control design and implementation are computationally expensive.

For these reasons, a simplified nonlinear airpath model of the diesel engine will be described

in Section 6.1. The model will be derived and parameterised for low and medium speed-load

points, which are covered by the New European Drive Cycle (NEDC). It will then be converted into

quasi-LPV form in Section 6.2. Finally, Section 6.3 will give some model validation results and a

comparison to the nonlinear model from Chapter 3.

A simplified version of the diesel engine model from Chapter 3 was suggested by Jankovic et al. for

nonlinear control design in [JJK98]. They propose a third order nonlinear model in which the

manifold dynamics are described solely by differentiating the ideal gas law pV = RT m resulting

in one differential equation for each the intake and the exhaust manifold pressure. The turbocharger

dynamics are approximated by the power transfer with time constant :

RTi Ti

pi = (Wci + Wxi Wie ) + pi

Vi Ti

RTx Tx

p x = (Wie + W f Wxi Wxt ) + px (6.4)

Vx Tx

1

Pc = (Pc + m Pt )

Subsequently, the mechanical efficiency of the turbocharger m will be set to unity as for the model in

Chapter 3. In a further approximation, the intake and exhaust manifold temperatures are assumed to

be constant such that the effect of Ti and Tx on pi and px , respectively, is neglected. Alternatively, the

derivatives could be neglected, but the measured temperatures could be kept as measured parameters

(quasi-static approach) to improve the model. However, with LPV control design in mind, the number

of parameters should be as small as possible, hence the choice to set these temperatures to constants.

Since the quasi-LPV model will be used for a whole range of speeds and loads (i. e. those covered on

the NEDC), this assumption limits the accuracy of the model and the constant temperatures will be

optimised to give a good representation of the airpath dynamics over the NEDC as discussed below.

The precompressor and postturbine pressures are fixed at ambient conditions. The relation

between compressor flow and power is given by (3.6) and stated here again for convenience:

c Pc

Wci = . (6.5)

c p Ta i

p

1

pa

6.1. Simplified Nonlinear Model 91

1

Note that the postcompressor pressure pc in (3.6) is identical to pi in this model and :=

=

0.286. Moreover, the compressor efficiency is now assumed to be constant rather than being param-

eterised as a function of turbocharger speed and pressure ratio across the compressor.

The flow through the EGR valve from (3.14) is approximated for subsonic conditions as suggested

in [GA98] by

Ar (xr ) px pi pi

Wxi = 2 1 , (6.6)

RTx px px

where the effective area of the EGR valve Ar is a quadratic polynomial in the valve lift xr . In order to

avoid the input xr entering the equations quadratically, the static nonlinearity from actuator position

to effective area is pulled out of the model and the effective area Ar is used as input directly. The

left-hand plot in Figure 6.1 shows that this static nonlinearity is monotonically increasing and can

therefore be inverted such that controllers can be designed with the effective area as output. The

right-hand graph in Figure 6.1 shows a plot of the non-dimensional flow function, i. e. the root term

2 1/

in (6.6), and its approximation. For pi

px

+1

0.53 the flow becomes choked, a condition

which is not met on the New European Drive Cycle (NEDC) and hence not considered here.

4

x 10 EGR Static Nonlinearity Flow Function Approximation

1.8 0.8

1.6 0.7

1.4

0.6

Nondimensional Flow

Effective Area in m2

1.2

0.5

1

0.4

0.8

0.3

0.6

0.2

0.4

0.2 0.1

original

approx

0 0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 0.5 0.55 0.6 0.65 0.7 0.75 0.8 0.85 0.9 0.95 1

Normalised EGR Position p /p

out in

Figure 6.1: Static nonlinearity of the EGR valve (left) and flow function approximation (right).

The mass flow rate from the intake manifold into the cylinders is determined by the speed-density

equation as for the model in Chapter 3:

pi N Vd

Wie = v (6.7)

Ti R 60 2

with the total displacement volume Vd and the volumetric efficiency v assumed to be constant for

this simplified model.

The turbine flow Wxt is also parameterised based on the standard orifice flow equation. In

order to allow for the conversion to a quasi-LPV model, the effective area is identified as a linear

function of the VGT position as opposed to the quadratic function in Section 3.4.3. Applying the

92 Chapter 6. Linear Parameter-Varying Modelling for Diesel Engines

same approximation for the flow function as in (6.6) the following parameterisation with identified

parameters a, b, c, d is obtained:

px px Tre f pa pa

Wxt = (axv + b) c 1 +d 2 1 , (6.8)

pa pre f Tx px px

where the reference temperature and pressure are the same as in Section 3.3, i. e. 298 K,101.3 kPa.

Figure 6.2 shows the fit (black) compared to experimental (red) and provided (green) data. Note

that the fit was chosen to match the experimental data especially at the low pressure ratios which are

encountered on the NEDC.

450

meas

400 prov

fit

350 0%

Corrected Air Flow in kg/h

300

20%

250

40%

200

60%

150

80%

100

90%

50

0

1 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2

Pressure Ratio p /p

out in

Figure 6.2: Turbine flow map for the quasi-LPV model: Measured, provided, and fitted data for

different VGT positions.

Finally, the turbine power is given in (3.7) and also stated here once more (note that the turbine

efficiency is constant in the simplified model):

pa

Pt = Wxt c p Tx t 1 . (6.9)

px

As opposed to the model derived in Chapter 3, the simplified version contains several constant

parameters, namely the compressor and turbine efficiencies c and t , the volumetric efficiency v ,

the intake and exhaust manifold temperatures Ti and Tx , and the time constant of the turbocharger

power transfer . Obviously, these parameters vary with engine conditions and keeping them constant

is only a crude approximation. However, it turns out that they capture the dynamics of the system

at least in the low and medium speed-load region of the NEDC, which is under investigation in this

project.

6.2. Conversion to a Quasi-LPV Model 93

Table 6.1: Parameter values of the simplified model resulting of a nonlinear optimisation.

c in % t in % v in % in s Ti in K Tx in K

61 76 87 0.11 313 509

The aforementioned parameters are chosen based on a nonlinear optimisation. Therefore, the

model is simulated with the inputs taken from the extra-urban part of the NEDC, which covers the

range of interest in this application (the reader is referred to Appendix D for more details). Matlab

is then used to find the parameters which result in the best fit of the simulation and experimental

data. As the optimisation is based on a nonlinear search, convergence to the global minimum cannot

be guaranteed. However, the obtained parameter values, which are given in Table 6.1, appear to be

reasonable.

The diesel engine model has a singularity at pi = pa when the compressor flow (6.5) becomes

infinite. Fortunately, it can be shown that the set := {( pi , px , Pc ) : pi > pa , px > pa , Pc > 0}

is invariant, i. e. every trajectory starting in stays in for all time [JJK98].

Note that the simplified third order nonlinear model has the same inputs and outputs as the higher

order airpath model (cf. Figure 4.1) used for control design in the previous chapter except with the

EGR input now being the effective area Ar instead of the actuator position xr . Engine speed and fuel

flow can again be considered as external inputs. However, they are not fixed parameters anymore,

but should now be regarded as disturbance inputs instead. This will be pointed out in more detail

subsequently.

Combining the equations from the previous section, the following three nonlinear differential equa-

tions are obtained:

RTi c Pc pi v Vd RTi px pi pi

pi = N+ 2 1 Ar

Vi c p Ta pi 1 Vi 2 60 Vi RTx px px

pa

Tx pi v Vd RTx px pi pi

p x = N 2 1 Ar

Ti Vx 2 60 Vx RTx px px

RTx px pa pa px Tre f

a c 1 +d 2 1 xv (6.10)

Vx pa px px pre f Tx

RTx px pa pa px Tre f RTx

b c 1 +d 2 1 + Wf

Vx pa px px pre f Tx Vx

Pc t c p Tx pa px Tre f px pa pa

Pc = + 1 a c 1 +d 2 1 xv

px pre f Tx pa px px

t c p Tx pa px Tre f px pa pa

+ 1 b c 1 +d 2 1

px pre f Tx pa px px

94 Chapter 6. Linear Parameter-Varying Modelling for Diesel Engines

These equations now have to be cast into the form (6.1). With the chosen states and inputs, the state

equation becomes:

Ar

pi pi

xv

p x = A((t)) px + B((t))

N

.

(6.11)

Pc Pc

Wf

An inspection of (6.10) reveals that only the intake and exhaust manifold pressure pi and px enter

the equations in various nonlinear ways. Therefore, they have to be considered as parameters,

i. e. 1 = pi and 2 = px . With that in mind, each term in (6.10) contains a state or input variable

plotted in boldface. The terms multiplying these variables now only contain fixed parameters or the

time-varying parameters 1 and 2 . Note that in some terms px is printed in boldface, which means

that there are no other states or inputs multiplying it and hence it can be left as a state.

In order to convert (6.10) to (6.11), all non-boldface printed pi and px are replaced by 1 and

2 , respectively, the terms multiplying states and inputs are separated, and everything is directly

written as a matrix equation. In (6.11), the origin is an equilibrium point, but it does not correspond

to a physical equilibrium point which will have intake and exhaust manifold pressures larger than

100 kPa. However, for the control design, the parameters are considered to be independent of

the state x, i. e. the parameter space will be gridded. Assuming there exist equilibria xeq and u eq

for each parameter value, (6.11) can be redefined by subtracting the equilibria from the states and

inputs. However, this does not affect the A and B matrices. In the implementation, the input to the

controller is the error between reference and measured values (i. e., operating point offsets disappear

in the controller input) and the controller will have integrating behaviour (such that the operating

point offsets in u can be provided by the integrators). Hence, a transformation of (6.11) to a physical

equilibrium is not necessary.

The significance of pi not appearing in boldface in (6.10) is that the first column of the A matrix

will be all zeros resulting in the system containing an integrator. This can be avoided if the speed

input N is redefined, e. g. N = N + 1500 rpm. Thus, the term multiplying N in the first row can be

restated as:

pi v Vd pi v Vd pi v Vd

N= N 1500. (6.12)

Vi 2 60 Vi 2 60 Vi 2 60

Now, the (1,1) entry in the A matrix becomes nonzero. The same applies to the second line in (6.10).

In a further step, the EGR and VGT inputs are redefined as

xv = xv + 50 %,

where the offsets correspond to the centre of the actuator range. Such a redefinition of the fuelling

input W f would result in a constant term not multiplying any other state or input; it is hence left

unchanged. Since the engine dynamics are not included in the model, zero fuel flow does not imply

6.3. Model Validation 95

engine stalling; the fuel flow is only included because it contributes to the flow from the cylinders

into the exhaust. Moreover, on the normal operating range of the diesel engine, the fuel flow is 18 to

70 times smaller than the air flow. Hence, assuming a nontypical zero fuel flow is not problematic.

Note that the more general case, in which (6.10) contains terms, i. e. constants or even nonlinear

functions in the scheduling parameters, that do not multiply any state or input can be converted into

the form (6.11) using Shammas state transformation [SC93].

The output equation in (6.1) can easily be obtained as

c 1 pi

0 0

Wci

=

c p Ta 1 1

pa

px (6.14)

pi 1 0 0 Pc

where pi is replaced by 1 and the D matrix is entirely zero. All tildes will be omitted subsequently.

Note that Ar and xv are the manipulated variables for this airpath submodel, while N and Wf are

external disturbances which can be measured but not be manipulated. Hence, N and Wf could be

introduced as additional parameters. For example, with N as parameter, pi could be used as a state

(described by a differential equation) rather than as an external parameter (which, by definition, can

vary independently of the state pi ) in two more places. However, in order to keep the computational

power required for the LPV control design small, N and W f are treated as external disturbances.

The non-uniqueness of the LPV representation of the nonlinear equations in (6.10) can make the

results conservative. This issue of the state-dependent representation can be addressed by including

an additional degree of freedom in the optimisation [HJ99]. Here, the chosen representation yielded

satisfactory performance (cf. Chapter 7) such that its optimisation was not necessary.

Note that there is no approximation involved when going from the simplified nonlinear model to

the quasi-LPV model. The validation in the following section therefore applies to both the simplified

nonlinear and the quasi-LPV model.

In order to assess the accuracy of the simplified nonlinear model, it will be compared to the validation

data provided for the full nonlinear model in Chapters 3 and 4 with a focus on the MAF (Wci ) and

MAP ( pi ) predictions. Figure 6.3 shows the simulation results for the same part of the NEDC as in

Figure 3.9.

The comparison to the full nonlinear model shows that the steady-state offset is larger in both

MAF and MAP for the quasi-LPV model, especially during the first ten seconds, but the latter

captures the transient behaviour rather well considering the crude approximations involved in the

derivation. Interestingly, during the transient between 25 and 35 seconds, MAP is better matched

than by the full nonlinear model, which shows a large overshoot. However, this occurs at the expense

of an underestimation of MAF. On the contrary, MAF is well matched during the first ten seconds,

where MAP is overestimated by the quasi-LPV model. It should be pointed out once more that

steady-state offsets are not of concern since any reasonably designed controller will take care of that.

96 Chapter 6. Linear Parameter-Varying Modelling for Diesel Engines

150

LPV

NL

140 Meas

MAP in kPa

130

120

110

100

0 5 10 15 20 25 30 35 40 45 50

200

150

MAF in kg/h

100

50

0

0 5 10 15 20 25 30 35 40 45 50

t in sec

Figure 6.3: Validation results for the quasi-LPV model (magenta) on the NEDC in comparison to

the experimental data (red) and the simulation results for the full nonlinear model (blue).

Note that the simulation is entirely open loop and that speed and load vary significantly as can

be seen in Figure 3.9. The EGR valve is again kept shut for this experiment to separate the effect of

the EGR path from the VGT path. The parameter estimation for the quasi-LPV model was obtained

from the complete extra-urban part of the NEDC (i. e. 400 s) with actuated EGR valve.

The further validation will focus on the fixed operating point chosen for the control design in the

previous chapter, namely 1500 rpm, 85 Nm. Figure 6.4 shows the response of the quasi-LPV model

to step inputs in EGR and VGT. The setup is identical to the one for Figure 4.2.

The gains in the main couplings are not as well matched as for the full nonlinear model. While

the gain in the EGR to MAF channel is slightly underestimated, it is overestimated for the VGT to

MAP channel. Again, this is not surprising when comparing the parameterisation effort that went

into the full nonlinear model as described in detail in Chapter 3.

Concerning the cross-couplings, the gain in the EGR to MAP channel is actually better matched

for the quasi-LPV model than for the nonlinear model. Note that this underestimated gain was the

cause for unexpected overshoots in the control designs in Chapter 5. Moreover, the nonminimum

phase behaviour of this channel is also reflected in the quasi-LPV model although it is hard to see in

the figure. The cross-coupling from VGT to MAF shows some transient response but the steady-state

is almost unchanged. This can be explained by the sensitivity of that gain to the effective area of the

6.3. Model Validation 97

100 100

Actuator in % 50 50

position

duty cycle

0 0

0 5 10 15 20 25 0 5 10 15 20 25

MAF in kg/h

NL

LPV

90 90

80 80

70 70

0 5 10 15 20 25 0 5 10 15 20 25

120 120

MAP in kPa

115 115

110 110

105 105

0 5 10 15 20 25 0 5 10 15 20 25

t in sec t in sec

Figure 6.4: Comparison of the MAF and MAP response to step inputs in EGR and VGT; experimental

data (red), nonlinear model (blue), quasi-LPV model (magenta).

EGR valve, where the gain even undergoes a sign change for a sweep in the EGR position as pointed

out in Section 4.1.2.

In conclusion it can be said that the validation results for the quasi-LPV model give confidence

that it will be useful for the control design in the next chapter. The rather crude approximations,

e. g. assuming constant turbine and compressor efficiencies, facilitate the generation of a nonlin-

ear model that allows a conversion to LPV form. Nonlinear models, which do not allow such a

direct conversion, can typically be approximated by quasi-LPV models. However, to find suitable

approximations is a substantial task.

The next chapter will utilise the quasi-LPV model for robust gain scheduling. The simulation

and experimental results given therein will show that this model is indeed very useful despite its lack

of accuracy.

98 Chapter 6. Linear Parameter-Varying Modelling for Diesel Engines

7

Linear Parameter-Varying Control Design

This chapter will present an LPV control design for the coordinated EGR-VGT control based on the

quasi-LPV model derived in Chapter 6. As pointed out earlier, this control methodology allows to

design a gain scheduled controller with stability and robustness guarantees. The underlying theory

will be presented in Section 7.1, which is meant to give a brief overview rather than a thorough

introduction. The interested reader is referred to the references given therein.

Section 7.2 describes the LPV control design based on H loopshaping. The designed controller

is scheduled on the intake manifold pressure and is valid over the entire speed-load range covered by

the NEDC (cf. Appendix D). Two controller extensions are considered in Section 7.3: Firstly, the

design of a command prefilter to speed up the response and secondly a speed and fuelling disturbance

rejection based on the disturbance model which is already included in the quasi-LPV model from

the last chapter.

Finally, Section 7.4 gives a variety of experimental results which indicate the suitability of the

LPV approach and then lead to the conclusions in Section 7.5.

Note that this brief presentation of the underlying theory is along the lines of [AA98], a reference

which is recommended for further information. Assume that the control problem is set up in the

general framework as depicted in Figure A.6 and used throughout the LTI control design in Chapter

5. Hence, we have the following state-space representation of the generalised LPV plant:

x A((t)) Bw ((t)) Bu ((t)) x

z = C z ((t)) Dzw ((t)) Dzu ((t)) w (7.1)

y C y ((t)) D yw ((t)) 0 u

where the shaped plant is assumed to be strictly proper (Dyu = 0) and the parameter vector (t)

is again constrained by (6.2) and (6.3). The design problem is to find a dynamic LPV controller

K (, )

with the state-space equations

x K = A K (, )x

K + B K (, )y

(7.2)

u = C K (, )x

K + D K (, )y

99

100 Chapter 7. Linear Parameter-Varying Control Design

which ensures internal stability and a guaranteed L2 gain bound for the closed loop from the exoge-

nous inputs w to the controlled outputs z, i. e.

T T

z z d 2

w w d, T 0

0 0

and zero initial conditions.

The following theorem gives a sufficient condition for the existence of such a controller and how

it can be constructed:

Theorem 7.1 (LPV synthesis) Consider the LPV plant governed by (7.1), with param-

eter trajectories constrained by (6.2) and (6.3). There exists a controller (7.2) enforcing

internal stability and a bound on the L2 gain of the closed-loop system (7.1) and (7.2),

whenever there exist parameter-dependent symmetric matrices X and Y and a parameter-

dependent quadruple of state-space data (K, L , M, N) such that for all (, )

d

the following infinite-dimensional LMI problem holds (terms denoted by are induced by

symmetry and only refer to the preceding term when used in a summation):

X I

>0 (7.3)

I Y

X + X A + () + LC y + ()

K + A + Bu N C y Y + AY + () + Bu M + ()

(X Bw + L D yw ) (Bw + Bu N D yw ) I

<0

C z + Dzu N C y C z Y + Dzu M Dzw + Dzu N D yw I

If a solution exits, an LPV controller is readily obtained with the following two-step scheme

A K = U 1 (X Y +U V + K X ( A Bu NC y )Y LC y Y X Bu M)V

B K = U 1 (L X Bu N)

C K = (M NC y Y )V

DK = N

Note that since all variables are involved linearly, the constraints (7.3) constitute a system of linear

matrix inequalities (LMIs). An LMI is a matrix inequality of the form

m

F(x) := F0 + xi Fi > 0 (or 0),

i=1

7.2. LPV Loopshaping Design 101

where x := [x1 x2 , . . . xm ] m is the vector of decision variables and the matrices Fi are sym-

metric and given apriori. Analytic solutions generally do not exist, but there are powerful numerical

algorithms available to find a feasible solution satisfying the convex constraints on the decision vari-

ables. A lot of control problems can be reformulated in LMI form, e. g. the H norm minimisation

via the Bounded Real Lemma [ZDG96]. In fact, Theorem 7.1 uses an extension of this lemma.

Theorem 7.1 constitutes an infinite-dimensional LMI problem due to its dependence on (, )

ranging over d . The remedy for this is to grid the parameter space (d needs only be checked

at its extreme points qi and q i since appears affinely in the LMIs). To keep the computational

effort feasible, the minimisation will typically be done for a defined grid and the constraints for the

optimal will then be checked for a finer grid. If this check fails, the minimisation needs to be

repeated for a finer grid. Theoretically, the gridding has to be finer than a certain limit based on

Lipschitz constants of the nonlinear system [WYP+96]. However, this limit is not readily available

and might require impractically fine gridding. Hence, gridding remains an ad hoc approach.

Moreover, the functional dependence of X and Y on has to be decided as well as for the

parameters K, L , M, N. There is no systematic rule for this but a practical approach suggested in

[WYP+96] is to mimic the parameter dependence of the plant.

For the controller implementation, I XY needs to be well-conditioned in order to avoid the

inversion of ill-conditioned matrices U and V . However, I XY will be nearly singular if the

constraint (7.3) is saturated at the optimum. This can be avoided by maximising the minimal

eigenvalue of XY by including the following LMI into the synthesis LMIs for a suboptimal value of

and maximising t

X tI

> 0. (7.4)

tI Y

In [AA98], Apkarian and Adams also provide an alternative formulation of Theorem 7.1 using

the projection lemma to eliminate the controller variables K, L , M, N which makes this approach

less conservative (since no controller structure has to be assigned). While the projection approach

requires a smaller number of scalar variables to be optimised, the resulting controller equations are

significantly more complex. The authors therefore suggest to use the projection approach for the

control design iterations and to generate the final controller using Theorem 7.1 for implementation.

Note that the control design based on quasi-LPV models is optimistic in the sense that it assumes

that the endogenous parameters (states) actually stay in d , which is a substantial task to prove

from the nonlinear equations. On the other hand, the quasi-LPV design is pessimistic (i. e. the

optimal is conservative) because it allows for the endogenous parameters to change independently

of the states which is physically not possible.

The quasi-LPV model derived in Chapter 6 is parameterised on the intake as well as on the exhaust

manifold pressure. In order to define a suitable grid, Figure 7.1 shows transient intake and exhaust

102 Chapter 7. Linear Parameter-Varying Control Design

manifold pressure measurements obtained from the NEDC including the extra-urban part. The plot

indicates that the exhaust manifold pressure can occasionally be smaller than the intake manifold

pressure. This effect is not included in the quasi-LPV model since it would result in a reverse EGR

flow. An inspection of the data plotted versus time reveals that those conditions are only met for very

small time periods during decelerations when the fuel is cut off and emissions are of no concern.

Hence, this effect will be neglected subsequently.

190

180

170

160

p in kPa

150

140

x

130

120

110

100

100 110 120 130 140 150 160 170 180 190

pi in kPa

Figure 7.1: Transient intake and exhaust manifold pressure measurements from NEDC.

The operating regime bounded by the black lines in Figure 7.1 has to be gridded for the control

design as explained in the previous section. This results in a large number of gridding points, e. g. 450

points using an equidistant mesh with grid points every 2 kPa (during the design it turned that this is

the maximal grid size to achieve the same optimal value when checked for a finer mesh). Therefore,

the controller designed in this chapter will only be scheduled on the intake manifold pressure to keep

the computational burden for the control design and especially for the controller implementation

as low as possible. This has the further advantage that an additional measurement of the exhaust

manifold pressure is not needed and only the standard engine sensors are required. Intuitively, both

the intake and exhaust manifold pressure give a good indication of the engine operating point in

the sense that larger pressures correspond to higher speeds and loads. Therefore, the advantage of

scheduling on both parameters is presumably minor, but an investigation of this issue is suggested

as further work.

Throughout this chapter, the exhaust manifold pressure will be fixed at 2.5 kPa larger than the

intake manifold pressure (depicted as the red line in Figure 7.1). This is a typical steady-state

pressure difference over the engine observed during testing. The sometimes significant deviations

from this assumption especially during transient operation of the engine require a controller with

good robustness properties. It will turn out that this is provided by the LPV controller designed

subsequently.

7.2. LPV Loopshaping Design 103

7.2.1 Scaling

Similar to the H loopshaping approach in Chapter 5, the quasi-LPV model has to be scaled. This

time, the input and output scaling matrices are derived from the operating range over the NEDC

(cf. Appendix D):

8.9 105 0 1

0

B SC = , C SC = 150

1

.

0 50 0 60

Remember that the actuator range for the EGR valve is now the effective area in m2 rather than the

actuator position leading to the (1,1) entry in the BSC matrix. Moreover, the values in this matrix

correspond to half the actuator range symmetric around the middle point as in (6.13).

For the LPV loopshaping design, the effect of the engine speed and fuelling disturbance is neglected;

this corresponds to an operating point of 1500 rpm and zero fuelling, due to the redefinition of the

inputs in (6.12). Also, the ambient pressure has to be fixed as it has been chosen in the quasi-

LPV model for the precompressor and postturbine pressures. Throughout this chapter, the ambient

pressure is fixed at pa =100 kPa. An investigation of (6.11) reveals that the ambient pressure appears

in ratios due to the flow equation. This introduces a strong nonlinearity especially for pressure ratios

close to unity which occur around idle. Fortunately, the controller performance for idling is not

crucial; in most cases, at least the VGT actuator will be saturated anyway. This is the justification for

a control design for intake manifold pressures larger than 102 kPa, thereby avoiding unity pressure

ratios. As will be seen, the controller still works for smaller pressure ratios, i. e. in idle state of the

engine.

Figure 7.2 shows the frequency response of the scaled quasi-LPV model (here without the actuator

position control loops) for three different intake manifold pressures, namely 102 (solid), 130 (dash),

and 160 kPa (dot), while the exhaust manifold pressure is fixed at 2.5 kPa larger than the intake

manifold pressure. The plots indicate the large gain variations of the system which supports the need

for gain scheduling. For example, the gain in the EGR to MAP channel varies by a factor of almost

ten and again confirms the finding from Chapter 4 that for low speeds and loads, and hence low

pressures, EGR has a strong effect on MAP. While the gain in the VGT channel appears to change

in linear correlation with the intake manifold pressure, the main gain change in the EGR channel

occurs at low pressures. The figure also supports the fact that the plant can be more easily decoupled

at higher pressures (with the appropriate relative scaling of the EGR and VGT channel).

As for the H loopshaping design in Chapter 5, the next step is to shape the plant including

the identified closed-loop transfer functions for the actuator position control loops (4.1) with a

precompensator containing integrators in both the EGR and VGT channel for zero steady-state error.

Due to the fast natural roll-off, a postcompensator is again not needed. The significant gain variations

of the plant over the operating envelope suggests that the precompensator should be scheduled on the

intake manifold pressure. The simplest approach is to only change the gain and leave the position

104 Chapter 7. Linear Parameter-Varying Control Design

EGR VGT

2 2

10 10

102 kPa

130 kPa

1 1

10 160 kPa 10

MAF

0 0

10 10

1 1

10 10

2

10

0 2 2 0 2

10 10 10 10 10

2 2

10 10

1 1

10 10

MAP

0 0

10 10

1 1

10 10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10

f in Hz f in Hz

Figure 7.2: Frequency responses of the scaled quasi-LPV model for different intake manifold

pressures (exhaust manifold pressure is 2.5 kPa larger than intake manifold pressure).

K EG R 1s s+20.4 0

W1 = 20.4 .

0 K V GT 1s s+20.2

20.2

In order to determine appropriate values for KEG R and K V GT , LTI loopshaping designs have been

carried out using the MATLAB command ncfsyn and the gains are chosen to achieve closed-loop

bandwidths of approximately 0.4 Hz for the MAF channel and 0.2 Hz for the MAP channel, which

justifies keeping the zeros fixed at these frequencies. Again, this choice is based on the experiences

from the control design in Chapter 5: Although the control task becomes easier at higher pressures,

the bandwidth cannot be increased due to the inner actuator position control loops. Table 7.1 gives

an overview of the chosen values for KEG R and K V GT as a function of MAP and the achieved LTI

stability margins.

Table 7.1: Precompensator gains and achieved LTI robust stability margins for different intake

manifold pressures.

MAP in kPa 102 105 110 120 130 140 150 160

K EG R 7.8 12.0 12.0 12.0 12.0 12.0 12.0 12.0

K V GT 9.0 6.0 6.0 4.0 3.3 2.5 2.0 1.5

bopt 0.43 0.50 0.51 0.54 0.55 0.56 0.57 0.59

Corresponding to Figure 7.2, the gain in the EGR channel only needs to be adjusted at low

pressures, while the gain in the VGT channel changes over the whole envelope. Moreover, the

7.2. LPV Loopshaping Design 105

increasing robust stability margin suggests that the control task becomes easier at higher pressures.

Note that for grid points for the LPV design between the provided MAP values in Table 7.1, the

gains will be interpolated.

In order to apply Theorem 7.1 for the control design, the dependence of X, Y , and the controller

matrices (K, L , M, N) on the scheduling parameter has to be decided. As mentioned earlier, the

guideline for this choice is to mimic the parameter dependence of the plant. Therefore, the entries

of the A matrix of the quasi-LPV model (6.11) are plotted in Figure 7.3 (the B and C matrix have

similar parameter dependences).

2 0.8 6

0.75

3 4

0.7

4 2

0.65

5 0.6 0

100 120 140 160 100 120 140 160 100 120 140 160

A21 A22 A23

7 20 1

0.5

6 40

0

5 60

0.5

4 80 1

100 120 140 160 100 120 140 160 100 120 140 160

A31 A32 A33

1 400 8

0.5 300

9

0 200

10

0.5 100

1 0 11

100 120 140 160 100 120 140 160 100 120 140 160

The plots in Figure 7.3 suggest that an affine parameter dependence is not sufficient, especially

considering the A(1,3) entry. An inspection of (6.10) shows that this is due to 1 entering the equation

as an inverse. Therefore, the following parameterisation achieving opt = 3.64 has been chosen:

1

X = X 0 + 1 X 1 + X2

1

and accordingly for the other matrices. The comparison of the optimal values for different param-

eterisations in Table 7.2 indicates that the guideline to mimic the plants parameter dependence is

actually justified.

Testing different grid sizes, it turns out that a mesh width of 2 kPa is sufficient for the intake

manifold pressure, i. e. 1 . At each grid point, the exhaust manifold pressure (i. e. 2 ) is fixed at

106 Chapter 7. Linear Parameter-Varying Control Design

Parameterisation X 0 + 1 X 1 X 0 + 12 X 1 X 0 + 1 X 1 + 12 X 2 X 0 + 1 X 1 + 1/1 X 2

LPV opt 13.84 8.14 4.00 3.64

2.5 kPa larger than the intake manifold pressure (2 = 1 + 2.5 kPa). It remains to choose the bound

on the rate of change of the parameter 1 . This can easily be determined from measurements and it

turns out that the maximal observed rate of change is less than 20 kPa/s.

For the controller synthesis, the LMI system (7.3) has been implemented with the Matlab LMI

Toolbox [GNL+95]. It takes around five hours of computing time on a 2 GHz Pentium IV until the

controller has solved the 90 LMIs with 412 decision variables. As mentioned before, the optimal

value is 3.64. To improve the conditioning of I XY , a 20 % suboptimal value has been chosen

(which results in bopt = 0.23) and the optimisation (7.4) is carried out, which takes about the same

time as the original controller synthesis. This optimisation improves the condition number of I XY

from 1.7 1019 to 1.0 1014 (evaluated at 1 = 102 kPa, 1 = 0 kPa/s).

Note that the optimal controller is of the same order as the generalised plant (i. e. seventh order,

three states from the quasi-LPV model, two from the actuator position control loop and another two

from the precompensator).

Figure 7.4 shows the magnitude plots and singular values of the synthesised controller for different

parameter values of the intake manifold pressure (1 = 0, 2 = 1 + 2.5). Although the gain of the

precompensator was varied for the design to compensate for the gain changes in the plant, the gains

of the controller still change significantly between the different operating points, especially in the

cross-couplings.

7.2.6 Implementation

The implementation of the controller requires the online computation of the controller in each time

step according to the equations provided in Theorem 7.1. With U chosen as identity matrix and V

as I XY , one matrix inversion (seventh order) has to be carried out each time step. An additional

matrix inversion is due to discretisation using Tustins transformation that requires the inverse of

I A K , which is also of seventh order. As for the LTI controllers in Chapter 5, the LPV controller

is implemented with the ECU sampling time of 16 ms and the Hanus anti-windup scheme is used for

the precompensator.

The state-space matrices of the generalised plant, which are also evaluated in each time step, are

stored as a map for 16 different values of 1 (2 is not involved in this step) and the matrices are

interpolated online. The rate of change of the intake manifold pressure is generated by numerically

differentiating the measurement after low-pass filtering it to reduce the noise.

7.2. LPV Loopshaping Design 107

MAF MAP 1

10

102 kPa

130 kPa

0 0

10 10 160 kPa

EGR

0

10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10 1

10

0 0

10 10

2

10

VGT

2 2

10 10

3

10 2 1 0 1 2

2 0 2 2 0 2

10 10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz f in Hz

Figure 7.4: Magnitude plots (left) and singular values (right) of the LPV controller for different

intake manifold pressures (1 = 0).

The closed-loop response of the control system is shown in Figure 7.5 again for the same three

different values of 1 . Although there are some differences in the bandwidth of the system, the

controller does a good job at providing similar closed-loop responses considering the significant

gain changes in the plant as depicted in Figure 7.2. This is especially true considering that the

responses are evaluated right at the boundaries of the operating range (i. e. 102 and 160 kPa). For

example, the dip in the closed-loop response of the MAF reference to MAF channel for 102 kPa

disappears when evaluated at 103 kPa.

10

102 kPa

130 kPa

0 0

10 10 160 kPa

MAF

0

10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10 1

10

0 0

10 10

2

10

MAP

2 2

10 10

3

10 2 1 0 1 2

2 0 2 2 0 2

10 10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz f in Hz

Figure 7.5: Magnitude plots (left) and singular values (right) of the closed loop with the LPV

controller for different intake manifold pressures (1 = 0).

108 Chapter 7. Linear Parameter-Varying Control Design

Regarding the MAP reference to MAP channel, the achieved bandwidth with the LPV controller

for 102 kPa (solid) is 0.4 Hz and hence twice as large as targeted and as achieved with an LTI controller

designed at this operating point with the same weights. As opposed to the reduced bandwidth in the

MAF channel mentioned before, this effect is actually observed for intake manifold pressures up to

105 kPa.

110 110

+FF

100 +Pre 100

LPV

90 Ref 90

MAF in kg/h

80 80

70 70

60 60

50 50

0 5 10 15 20 25 30 0 5 10 15 20 25 30

120 120

115 115

MAP in kPa

110 110

105 105

0 5 10 15 20 25 30 0 5 10 15 20 25 30

t in sec

t in sec

Figure 7.6: Validation of the LPV loopshaping controller with steps in the MAF and MAP references

without feedforward (red), with H feedforward compensator (blue), and with a command prefilter

(green).

Figure 7.6 (red traces) gives a first impression of the controller performance by depicting the

response to step changes in the MAF and MAP reference as they were used in Chapter 5. Comparing

these results to the LTI H loopshaping controller in Figure 5.5 reveals a similar performance at

this particular operating point except for a slight overshoot in the MAF response and a slightly

slower response to the MAP step. This can be explained by investigating Figure 6.4 which shows

that opposed to the full nonlinear model the LPV model underestimates the gain in the MAF and

overestimates the gain in the MAP main coupling. Considering that this controller is based on a

simplified nonlinear model and designed for a wide operating range, the achieved performance is

rather impressive. Further validation results will be provided in Section 7.4 from experiments with

the NEDC, i. e. including significant changes in speed and load as well as in the MAF and MAP

reference. As in Chapter 5, the experimental data have been low-pass filtered with a cut-off frequency

of 1.5 Hz for the presentation of the results unless otherwise stated.

7.3. Controller Extensions 109

Two controller extensions will be discussed in this section. Firstly, a feedforward extension of the

LPV controller using a model matching approach as in Chapter 5. Secondly, a disturbance rejection

for speed and fuelling disturbances since the LPV controller operates on a wide speed and fuel range

rather than at a fixed operating point. Potentially, these controller extensions could be incorporated

into the generalised plant for the LPV synthesis. However, this increases the computational burden

significantly and a modular approach is taken here. Both the feedforward and the disturbance rejection

will be designed for the closed-loop system (i. e. including the LPV controller) at a fixed operating

point. This approach relies on the LPV controller rendering the closed-loop transfer functions from

the input and disturbance to the output similar at each operating point, cf. Figure 7.5.

7.3.1 Feedforward

In order to design a feedforward extension to the LPV controller, an LTI H model matching problem

will be solved. While the two degrees of freedom design in Chapter 5 simultaneously synthesised the

feedback and feedforward controllers, the approach pursued here is sequential. Based on the setup

depicted in Figure 7.7, where the reference model is chosen as in (5.2) and KL PV is the LPV controller

for 1 = 111 kPa, = 0 kPa/s, the feedforward element Kr is designed by minimising the H norm

from w to (u, e). The weight Wr consists of a first order filter to especially penalise deviations from

the reference model in the low frequency range of interest. To achieve zero steady-state error, the

steady-state gain of Kr is inverted and included in the implementation.

,

Plant + Act. Pos. Control

w r u y e

Kr K L PV W1 B SC Pact Peng C SC Wr

Tr e f

The results are also shown in Figure 7.6. Although difficult to see on that scale, the 90 % rise

time in the main channels has been reduced by 0.5 s for the LPV controller with feedforward (blue

traces). Moreover, the cross-coupling from MAF to MAP is significantly reduced. As opposed to the

two degrees of freedom controller from Chapter 5, where a gain underestimation in the full nonlinear

model lead to an overshoot in the MAP reference to MAP channel (Figure 5.10), the aforementioned

gain overestimation in the LPV model leads to a more damped response although the reference

models are the same and in both cases well matched for the closed loop.

110 Chapter 7. Linear Parameter-Varying Control Design

Alternatively, instead of setting up a model matching problem, a simple command prefilter can

be used to overdrive the reference and therefore speed up the response of the control system. The

idea is to consider the closed-loop as a first order system with a bandwidth of 0.4 Hz and 0.2 Hz for

the MAF and MAP channel, respectively. Note that those were the intended bandwidths, but the

achieved bandwidths might be somewhat different depending on the operating point as indicated in

Figure 7.5. The command prefilter is then chosen as a simple lead-lag filter with the zero cancelling

the pole of the closed loop and in turn introducing a faster pole:

1/0.4s + 2 1/0.2s + 2

TEG R = , TV GT = .

1/0.8s + 2 1/0.4s + 2

Figure 7.6 also depicts the experimental results with the command prefilter (green). While the MAF

Reference to MAF channel achieves a performance very similar to the H feedforward compensator,

both cross-couplings show a bigger tracking error. The MAP main response initially rises faster, but

does not reach the reference until the integral part kicks in. Due to the initially faster response, less

error builds up in the integrator which leads to the effect, that the 90 % rise time is actually larger

than without prefilter.

88

+FF

87 +Pre

MAF in kg/h

LPV

86

85

84

83

9 9.5 10 10.5 11 11.5 12

130

125

MAP in kPa

120

115

110

105

9 9.5 10 10.5 11 11.5 12

t in sec

Figure 7.8: Modified references with feedforward (blue), command prefilter (green), and without

feedforward (red).

The responses of the H feedforward compensator (blue) and command prefilter (green) are

plotted in Figure 7.8 for comparison. The red traces are for the LPV controller without feedforward

and therefore identical to the reference, which is chosen as the step in the MAP reference from Figure

7.6. The advantage of the H feedforward compensator is that it also changes the MAF reference

to reduce the cross-coupling. Moreover, the overdrive in the MAP reference is smaller, but better

matched to the system as revealed in Figure 7.6.

Until now, the effect of speed and load on MAF and MAP has been ignored. Changes in these

variables will affect MAF and MAP creating a tracking error which will then be corrected by the

7.3. Controller Extensions 111

LPV controller. This sort of disturbance rejection is hence inherent in the LPV control system.

Since the disturbance can be measured, it is likely that this response can be improved. As mentioned

before, it would be possible to extend the LPV optimisation to a generalised plant including the

disturbance model (note that this model is provided by the quasi-LPV model already, it is simply the

transfer function from N and W f to the outputs). However, this would result in a significant increase

in complexity and hence required computing power.

An alternative approach is to use the fact that the closed-loop system with the LPV controller

also renders the transfer function from the disturbance to the output similar for a wide range of the

operating envelope. Therefore, for the operating point 1 = 111 kPa and 1 = 0 kPa/s, a feedforward

element Kd is designed using the setup depicted in Figure 7.9.

d

D SC Wd

, Kd Pd

Plant + Act. Pos. Control

ud

r y

K L PV W1 B SC Pact Peng C SC

The LPV controller K L PV is included and the design synthesises the controller Kd minimising

the transfer function from d to (ud , y). The disturbances are scaled by their range over the NEDC

(N in rpm and W f in g/s) leading to the scaling matrix

2500 0

D SC = .

0 3

The weighting matrix Wd is introduced to ensure that the steady-state gain from d to y is close to

zero. As for the model matching in Section 5.2, the H optimisation only minimises the maximal

singular value. Hence, after some design iterations, the following weight has been chosen

2 104 s+20.001

s+210

0

Wd = .

0 1 104 s+20.001

s+210

Figure 7.10 shows the improvement in the disturbance rejection by comparing the closed-loop transfer

functions from d to y with (solid) and without Kd (dash) at the chosen operating point 1 = 111 kPa

with the design weight removed but included scaling DSC . Up to a frequency of 1 Hz the improvement

is significant in both the speed and the fuelling disturbance channel. Beyond 1 Hz, the transfer

functions are almost identical, which is reasonable considering the limited bandwidth of the actuator

position control loops.

112 Chapter 7. Linear Parameter-Varying Control Design

N Wf 1

10

LPV

+DR

0 0

MAF 10 10

0

10

2 2

10 10

2 0 2 2 0 2

10 10 10 10 10 10 1

10

0 0

10 10

2

10

MAP

2 2

10 10

3

10 2 1 0 1 2

2 0 2 2 0 2

10 10 10 10 10 10 10 10 10 10 10

f in Hz f in Hz f in Hz

Figure 7.10: Magnitude plots (left) and singular values (right) of the closed-loop transfer function

from disturbances to outputs with Kd (solid) and without (dash); 1 = 111 kPa, = 0 kPa/s.

Experimentally, the disturbance rejection is evaluated for constant MAP and MAF reference

values but steps in the disturbance channels, i. e. engine speed and fuelling. Note that this is a severe

test since the MAP and MAF references typically change with speed and load. Figure 7.11 shows

the results. The upper plots show the disturbances, the middle plots the MAF and MAP responses,

and the lower plots the actuator positions.

The disturbance rejection performance has significantly increased in both channels. Due to

the fact that the engine speed almost immediately affects the flow into the engine by changing the

pumping rate, while the pressure needs some time to build up, the performance improvement is more

obvious in the slower MAP response. Since the disturbance rejection controller Kd improves the

performance up to a frequency of around 1 Hz as depicted in Figure 7.10, the initial deviation from

the MAF and MAP setpoint after each disturbance step change is not reduced as expected.

So far, the LPV controller has been tested on somewhat artificial conditions, e. g. step changes in

the MAP and MAF references for fixed speed and load. This enables a better comparison to the

H controllers from Chapter 5, which were designed for a fixed operating condition. The LPV

approach, however, allows the controller to function on a range of speeds and loads since changes in

these conditions will result in a change of the intake manifold pressure, i. e. the scheduling variable.

A more realistic scenario for the LPV controller evaluation is provided by the NEDC which means

simultaneous changes in speed and load as well as in the MAP and MAF references (which are

actually mapped based on speed an fuelling). Figure 7.12 shows some of the results taken from the

extra-urban part of the NEDC.

The upper plots in Figure 7.12 depict the speed and load traces and show that this part of the

NEDC covers the whole range for which the LPV controller was designed. The red traces in the

7.4. Experimental Results from NEDC 113

1800 0.8

Wf in g/s

N in rpm

1600 0.7

1400 0.6

1200 0.5

0 10 20 30 40 0 10 20 30 40

110

LPV 114

MAF in kg/h

Ref

MAP in kPa

100

+DR

112

90

110

80

108

0 10 20 30 40 0 10 20 30 40

80 65

EGR Pos in %

VGT Pos in %

60 60

40 55

20 50

0 45

0 10 20 30 40 0 10 20 30 40

t in sec t in sec

Figure 7.11: Response to steps in the speed and fuelling disturbances for the LPV controller with

(red) and without (blue) disturbance rejection.

middle plots indicate that the LPV controller tracks the references (black) nicely. Since the LPV

controller is purely based on a rather simple model without tuning on the engine, this result is rather

promising. The green traces represent the LPV controller with disturbance rejection. While the

MAF tracking is almost identical, the MAP tracking is slightly slower (thereby reducing some slight

overshoots of the pure LPV controller). This surprising result is due to the fact that the disturbances

actually act in favour of the signals to be tracked, e. g. a speed increase leads to an increase in MAF

and MAP, which is desired.

Another interesting effect is that the LPV controller with feedforward element also yields a slower

response in the MAP channel. This can be explained as follows: In order to track an increasing

MAF reference, it is beneficial to open the VGT, thereby reducing the EGR flow and hence increasing

MAF. However, this reduces the tracking performance if the MAP reference increases at the same

time. This is exactly what happens with the feedforward compensator, cf. the VGT position trace

in the lower right plot in Figure 7.12 at around 10 s. The dip in the VGT position slows down the

MAP tracking, while the MAF tracking does not improve noticeably. The only beneficial effect of

the feedforward is that it reduces the slight MAP overshoot of the pure LPV controller at 80 s.

Engaging both the feedforward and disturbance rejection at the same time does not yield a

significant performance difference to the tests with each part individually; it is hence not plotted

here. Note that the engine strategy actually generates the MAF and MAP references via static maps

based on speed and load. Therefore, one could think of a control design which takes this correlation

(as well as the disturbance model) into account and in that framework design a feedforward from the

114 Chapter 7. Linear Parameter-Varying Control Design

150

2500

100

N in rpm

T in Nm

2000

50

1500

l

0

1000

50

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

160

200 Ref

MAF in kg/h

MAP in kPa

150 140

100

120

50

0 100

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

100 100

LPV

EGR Pos in %

VGT Pos in %

+FF

+DR

ECU

50 50

0 0

0 20 40 60 80 100 120 140 0 20 40 60 80 100 120 140

t in sec t in sec

1800

100

1700

N in rpm

T in Nm

1600

50

1500

l

1400 0

5 10 15 20 5 10 15 20

100 120

MAF in kg/h

115

MAP in kPa

80

110

60

105

Ref

40 100

5 10 15 20 5 10 15 20

100 100

LPV

EGR Pos in %

VGT Pos in %

+FF

+DR

50 ECU 50

0 0

5 10 15 20 5 10 15 20

t in sec t in sec

Figure 7.12: Performance of the LPV controller (red) with FF (blue) and with disturbance rejection

(green) in comparison to the ECU controller (magenta) on extra-urban part of the NEDC. Note that

the VGT position measurement was not available for the ECU controller tests.

7.4. Experimental Results from NEDC 115

MAF and MAP references to the actuator position control setpoints (an additional feedforward from

the disturbance is not required since the changes in speed and load are immediately reflected in a

change of the references). However, the static lookup tables would have to be incorporated into the

control design in a suitable way. Due to their high nonlinearity, this is a substantial task and beyond

the scope of this thesis.

A comparison to the fully calibrated ECU controller (magenta traces in Figure 7.12) shows that

the latter has a slightly worse MAP tracking performance, but the MAF tracking for pedal tip-ins

is significantly better than for the LPV controller (cf. magnified plots). This could be due to the

fact that the ECU controller has a feedforward path from speed and fuelling to the duty cycle of the

EVRV actuator directly, while the LPV controller even with feedforward only generates a position

reference for the underlying control loop which slows down the response. However, including

the feedforward from the ECU controller into the implemented LPV controller hardly yields any

performance improvement at all. Also, the LTI design in Chapter 5 (cf. Figure 5.20) has shown that

the ECU controller performance can be met without a direct feedforward to the EGR duty cycle. This

will be discussed further in Section 7.5. Note that a fast MAF response is important for driveability

since the amount of fuel that can be injected is smoke limited by the amount of air in the cylinders.

150

2000

100

N in rpm

T in Nm

1500 50

l

1000 0

50

0 10 20 30 40 0 10 20 30 40

130

Ref LPV

MAF in kg/h

Ref ECU

MAP in kPa

100 120

50 110

0 100

0 10 20 30 40 0 10 20 30 40

100 100

EGR Pos in %

VGT Pos in %

50 50

LPV

+FF

+DR

ECU

0 0

0 10 20 30 40 0 10 20 30 40

t in sec t in sec

Figure 7.13: Performance of the LPV controller (red) with FF (blue) and with disturbance rejection

(green) in comparison to the ECU controller (magenta) on urban part of the NEDC. Note that the

VGT position measurement was not available for the ECU controller tests.

For further evaluation, Figure 7.13 shows 40 s of the urban part of the NEDC with a more

aggressive driving pattern including gear changes. The MAF traces are again very similar for the

pure LPV controller (red) and including feedforward (blue) or disturbance rejection (green). For

116 Chapter 7. Linear Parameter-Varying Control Design

the MAP channel, the trend is similar to Figure 7.12 with the pure LPV controller tending to have

some overshoot in the MAP channel. A significant overshoot in MAP appears at around 8 s. This is

avoided by the slower LPV controller with disturbance rejection. Also, the LPV with feedforward

does not have this overshoot, the VGT position is reduced quicker (blue trace in the lower right plot),

presumably to support the further increase in MAF as discussed before.

A comparison to the ECU controller (magenta) shows again the quicker response in the MAF

channel. For this more aggressive driving pattern, this has also the following effect on the MAP

channel: Due to the faster increase in MAF, the smoke limited fuelling can be increased at a higher

rate which results in the MAP reference (being mapped on speed and fuelling), to increase faster.

This is plotted in the middle right plot (black, dash) and results in a faster MAP response for the

ECU controller compared to the LPV controllers. Another feature of the ECU controller is a gear

change strategy which keeps the MAF high (emissions are of less concern since the fuelling is low)

presumably to be able to inject more fuel directly after re-engaging. This slower MAF reduction

during deceleration can also be seen in Figure 7.12 at around 111 s.

7.5 Conclusions

In conclusion, it can be said that the implementation of the LPV controller was very successful, espe-

cially considering that the plant changes significantly with the engine operating point. Although the

design procedure is rather complex and the implementation of the controller imposes strong require-

ments on the hardware for real-time implementation, it provides a model-based design methodology

and avoids a lot of mapping and calibration effort. The main implementation problem with this

LPV design is that the controller has to be computed in each time-step depending on the scheduling

variable. This can potentially be avoided by computing controllers for a variety of possible param-

eter values, storing the data in maps, and using online interpolation. However, this might be very

sensitive depending on the state-space representation of the controller.

Although the underlying model is a rather crude approximation, the LPV controller works surpris-

ingly well. This indicates that the nonlinearities and dynamics of the plant have been well captured by

this approach. The comparison to the ECU controller reveals one shortcoming of the LPV controller

designed here, namely the slower response in the MAF channel in part of the operating regime. This

is not surprising since the LPV controller is purely model-based. It is therefore required to be able

to adjust the controller parameters online for optimal performance. This is easily achieved with PI

controllers (which makes them so popular in engine control), but not straightforward with more com-

plicated schemes like LPV control. A possible solution has been suggested in [Chr02], where it is

proposed to parameterise the weights with a parameter affecting their bandwidths or static gains. The

resulting controller is hence gain scheduled with parameters from the plant and with the parameters

introduced for calibrating the control system performance. However, this makes the computation

and implementation of such controllers more complex. Note that the calibration task with such an

extended control scheme is still small compared to the overall calibration of the EGR-VGT system

and hence does not contradict the goal to reduce the calibration effort by model-based design.

7.5. Conclusions 117

The robustness of the LPV controller implies that it should immediately work on similar engines

considering the crude approximations of the LPV model. However, even for a different engine, the

modelling effort is very small compared to the full nonlinear model from Chapter 3 requiring the

parameterisation of the compressor and turbine flow and efficiency maps. For the quasi-LPV model,

only the turbine and EGR flow maps have to be re-identified as well as the parameters in Table 6.1.

Finally, note that the LPV controller has been designed based on the model parameterised on the

NEDC with the operating ranges given in Appendix D. However, the controller has also been tested

successfully well beyond this limit, which is probably due to the fact that the airpath dynamics most

significantly change in the low speed and load region.

118 Chapter 7. Linear Parameter-Varying Control Design

8

Conclusions

Mean-value modelling and robust control of the airpath of a turbocharged diesel engine have been

investigated in this thesis. Since a detailed discussion of the results can be found in Section 5.7

and Section 7.5, it remains here to point out the main contributions of this dissertation and to give

suggestions for further work.

Turbocharger parameterisation for low and medium speed-load points. The parameterisation

of variable geometry turbochargers for mean-value modelling is typically based on compressor

and turbine flow and efficiency maps provided by the supplier. At low turbocharger speeds, and

hence low air flows, the heat exchange via the turbocharger housing affects the temperature-based

measurements of the efficiencies. Therefore, the low-speed operating regime of the turbocharger is

excluded from the supplied maps and mean-value models mainly rely on the extrapolation into this

region, which is met on emission drive cycles for most of the time, and hence of great significance.

Based on experimental data from the engine under investigation, it is shown that while the flow

maps extend from the high-speed region in a natural way the efficiency maps are severely affected

by heat transfer. It is argued that this effect should be included in the mean-value model. A physics-

based parameterisation is suggested for the turbine efficiency, which poses the biggest problems in

turbocharger modelling. The simulation results with the new model structure match the experimental

transient engine data rather well.

Uncertainty investigation of the airpath model. Based on the comparison between identified

and simulated frequency responses of the airpath model, the effect of parameter variations has

been investigated. It turns out that the parameters commonly perceived to be the most uncertain,

i. e. compressor and turbine efficiency as well as the effective area of the EGR valve, have the biggest

effect on the frequency response. Changes in these parameters can explain the observed differences

between experimental and simulation data.

parameterise the observed uncertainty in the airpath model are compared and evaluated. Based on an

extended H loopshaping procedure, two degrees of freedom controllers have been designed using

119

120 Chapter 8. Conclusions

synthesis tools for a general coprime factor uncertainty description, and for application tailored

additive and state-space uncertainty parameterisation. The designed controllers are compared using

real analysis and based on simulation and experimental data. As one would expect, the controllers

based on application tailored uncertainty outperform the conventional loopshaping controller.

Control-oriented linear parameter-varying airpath modelling. To allow for robust gain scheduled

control design, a third order nonlinear airpath model of the engine under investigation is derived,

which is then converted into linear parameter-varying (LPV) form with the intake and exhaust

manifold pressure as parameters without approximations involved. The model predicts the mass air

flow and manifold absolute pressure based on the EGR and VGT actuator position, speed, and load.

A comparison to experimental data from the engine test bed and to the higher order nonlinear model

suggests the validity of this approach.

Linear parameter-varying airpath control. Based on the quasi-LPV model of the airpath, a

robustly gain scheduled LPV controller is synthesised in an LMI framework and implemented on the

engine. The experimental results are very promising and indicate that this recently developed design

methodology is suitable for this control problem. It also shows that the quasi-LPV model captures

the nonlinearities and dynamics of the plant sufficiently well. Since the MAF tracking performance

of the fully calibrated ECU controller is not met by the LPV controller on parts of the operating

regime, it is suggested to take the LPV controller as a baseline controller. Possible extensions are

discussed in the next section.

The modelling and control investigations in this dissertation have shown that there is still a lot of

room for improvement of mean-value models of turbocharged diesel engines in general, especially

for simulation purposes. On the other hand, the LPV control design has shown that coordinated EGR-

VGT controllers can successfully be synthesised using a rather simple model. Since this technique

has only been developed recently, the suggestions for further work focus on this particular area.

Improvements of the quasi-LPV model. The LPV model derived in this dissertation assumes

constant intake and exhaust manifold temperatures. However, they could be included as measured

parameters while still neglecting the effect of their derivatives and hence not increasing the dynamic

order of the model (quasi-static approach). This extension might be especially useful when extending

the operating range of the model to high speed and load points. It should also be investigated as to how

far combustion and engine dynamics can be included into the model and what kind of approximations

have to be made to still retain the LPV form of the model.

straints on computing power for the design and the implementation, the LPV controller has been

chosen to be scheduled on the intake manifold pressure only. The other parameter of the quasi-LPV

8.2. Further Work 121

model, i. e. the exhaust manifold pressure, is kept fixed at 2.5 kPa larger than the intake manifold

pressure. However, there are significant pressure differences especially during transient operation.

This motivates an investigation of the benefits that can be gained by scheduling on the exhaust man-

ifold pressure as well. If manifold temperatures are included as parameters of the model, they can

also be used as additional scheduled variables.

achieve optimal tracking performance, it is suggested to make the LPV controller calibratable. One

possible solution has already been discussed, where the weights are parameterised by additional

scheduling parameters to allow for online tuning of the bandwidth or static gain of the weights.

However, this increases the computational load significantly and alternative solutions are desirable.

Even with extensions to allow for online tuning, the overall calibration task will still be reduced

compared to the current ECU controller. Moreover, some engine tuning is always necessary to

achieve optimal performance.

Implementation of the LPV controller on an ECU. In the test cell, the LPV controller is run in

real-time on a dSPACE system. Ultimately, the control scheme has to be coded into the ECU which

is much more constrained with regard to computing power and memory. It should therefore be

investigated how the implementation can be changed to accommodate for this purpose, e. g. off-line

evaluation of the controller for different parameter values and then storing the data as a map.

Alternative controllers based on the simplified model. For better assessment of the LPV controller

performance, it would be interesting to compare its performance to other control schemes based on

the same model. All the required data for the simplified nonlinear model is provided.

122 Chapter 8. Conclusions

A

Robust Control Theory

This appendix is intended to introduce the reader, who might not be familiar with robust control,

to the main concepts encountered in this thesis. For an introductory text on single-input single-

output (SISO) robust control theory the reader is referred to [DFT92]. An overview of multivariable

(MIMO) robust control theory is presented in [ZD98], which also contains a lot of problems for self-

study. A complete treatment of the topic as well as an excellent reference can be found in [ZDG96].

To facilitate the computer-aided control design, the -Analysis and Synthesis Toolbox for Matlab

has been developed, which is used throughout this report. The manual of this toolbox [BDG+94] is

also a very good source for a summary of the theory combined with extensive design examples.

A.1 Fundamentals

A.1.1 State-space Representations

The robust control design presented in this chapter assumes finite-dimensional linear time-invariant

(FDLTI) systems. They are conveniently described using the state-space description (particularly

for the MIMO case).

x(t) = Ax(t) + Bu(t)

y(t) = Cx(t) + Du(t)

describes a continuous-time system with the states x(t) n , the inputs u(t) m , and the outputs

y(t) p . Applying the Laplace transformation gives the transfer matrix from input u to output y

A B

G(s) = C(s I A)1 B + D =

C D

A B

where denotes a state-space realization.

C D

The way to deal with robust stability and performance in the framework of H robust control is

to minimise the H norm of certain transfer functions. Intuitively, if we minimise the norm of

123

124 Chapter A. Robust Control Theory

the (closed-loop) transfer function from the reference input to the error signal, we should expect

a good tracking. Similarly, disturbance rejection could be achieved by minimising the transfer

function from the disturbance to the output signal. In SISO systems, the H norm of a transfer

function corresponds to the maximum value of its frequency response. Minimising the H norm

therefore corresponds to a design with respect to a worst-case scenario, which is hence suitable for

a design with respect to robustness. For MIMO systems, the H norm is given by the maximum

over frequency of the largest singular value of the transfer matrix. This can be shown to be the least

upper bound of the energy of the output for unit energy input (indeed, the H norm is induced by

the input-output mapping of signals in H2 ). Therefore, the meaning of the H norm as a sort of

worst-case gain carries over to the multivariable case.

Consider the Banach space L of complex matrix functions that are (essentially) bounded on

the imaginary axis with the norm

G

:= ess sup [G( j )]

where denotes the largest singular value. The H space is a closed subspace of L with functions

that are analytic and bounded in the open right-half plane and the norm defined as

G

:= sup [G(s)] = sup [G( j )].

Re(s)>0

RH is the real rational subspace of H , which consists of all proper and real rational transfer

matrices. While the size of the transfer matrices is measured using the H norm, the signals

considered in this report are supposed to be in H2 . Theses are the signals, which are defined for

positive time and zero for negative time with a bounded L2 norm given by

x

2 := x (t)x(t)dt

Note that the step function does not belong to this group of signals with bounded energy, however,

it is energy bounded for any finite time.

SISO systems are typically characterised by a transfer function consisting of a numerator and de-

nominator polynomial in the Laplace variable s with real-valued coefficients. This concept extended

to MIMO systems leads to coprime factorisations, which represent transfer function matrices as the

product of two factors, i. e. G = M 1 N where M and N are left coprime over RH . Two matrices

M and N in RH are left coprime over RH if they have the same number of columns and if

there exist matrices Xl and Yl also in RH such that

Xl

M N = M X l + N Yl = I.

Yl

A.1. Fundamentals 125

Note that since the multiplication of matrices does not commute, a right coprime factorisation,

i. e. G = N M 1 can be introduced whose properties follow by duality. The idea of H loopshaping

in Section A.2.2 is based on the stabilisation of a perturbed coprime factor plant description.

The quality of a control design is usually assessed in terms of the following three categories:

1. Stability

3. Dynamic behaviour

However, before a control design can be assessed, it has to be made sure that the control loop is

well-posed. For example, consider the standard feedback configuration in Figure A.1, which is

basically the same block diagram as in Figure 1.3 augmented with additional disturbance and noise

signals. The plant P and the controller K are fixed real rational proper transfer matrices. This

feedback system is said to be well-posed if all closed-loop transfer matrices are well-defined and

proper, which is true if and only if I + K ()P() is invertible.

di d

r u up y

K P

n

The first requirement a well-posed control system has to satisfy is stability. To determine stability,

it is in general not sufficient to consider only the input-output transfer function of the closed-loop

Tyr in Figure A.1, because it is possible that internal signals blow up due to unstable closed-loop

transfer functions from any external signals (r, n, d, di ) to any internal signal (y, u, u p ). To avoid

this, the notion of internal stability is introduced, which guarantees that all signals in the closed-loop

system are bounded provided that the injected signals at any location are bounded (as well as the

initial conditions).

Definition A.1 (Internal Stability) The feedback system in Figure A.1 is internally stable

if and only if it is well-posed and the transfer matrix

1

I K (I + K P)1 K (I + P K )1

=

P I P(I + K P)1 (I + P K )1

belongs to RH .

126 Chapter A. Robust Control Theory

Note that if P RH and K RH Def. A.1 is equivalent to det(I + P(s)K (s)) having no zeros

in the closed right-half plane. Moreover, note that the transfer matrix in Definition A.1 corresponds

to the transfer matrix from (d1 , d2 ) to (e1 , e2 ) in Figure A.3 if is replaced by P and Tzw by K

(the minus sign is due to the positive feedback convention in that figure).

The essential feature of robust control is to explicitly consider uncertainties in the model of the plant.

Uncertainty describes the difference between the mathematical model and reality. A physical system

can never be modelled perfectly but one usually has an idea of a set of models, which contains the

real plant. This is for example the case when the parameters of a model are known to lie in certain

intervals. Moreover, it is sometimes necessary to simplify given models for control design, e. g. to

linearise a nonlinear plant. In this case, the neglected dynamics can be considered by including

them in the uncertainty. For robust control design the system is described as a nominal plant plus

uncertainty and the resulting robust controller is optimised for that set of plants.

There are several ways of describing uncertainty. For illustration, consider a multiplicative

uncertainty of the form

with the nominal model P(s) and the uncertainty described by the stable weighting matrices W1 (s),

W2 (s) and

(s)

< 1. This uncertainty structure can be represented in a block diagram as shown

in Figure A.2.

z w

W2 W1

r y

K P

For any given controller, the closed-loop internal stability for any plant P (where =

(I + W1 W2 )P is the set of plants defined by the nominal plant P and the multiplicative uncertainty)

can be determined using the following robust stability test:

stabilising controller for the nominal plant P. Then the closed-loop system is well-posed

and internally stable for all RH with

< 1 if and only if

W2 To W1

1.

A.1. Fundamentals 127

proving Theorem A.1 is to determine the transfer function from w to z and to apply the small gain

theorem:

Theorem A.2 (Small Gain Theorem) Let Tzw RH and > 0. Then the system in

Figure A.3 is well-posed and internally stable for all (s) RH with

< 1/

if and only if

Tzw

.

The small gain theorem is also very useful to derive robust stability conditions for other kinds of

uncertainties (additive, coprime, etc.).

d1 e1

e2 d2

Tzw

The control design in the subsequent chapters is based on the so-called general framework depicted in

Figure A.4. The general framework consists of the generalised plant P, the controller to be designed

(or to be analysed) K , and the uncertainty . The vector signal w contains the exogenous inputs

and the vector signal z the signals to be controlled. The generalised plant has the form

P11 (s) P12 (s) P13 (s)

P(s) = P21 (s) P22 (s) P23 (s)

P31 (s) P32 (s) P33 (s)

Note that any interconnected system such as the feedback configurations in Figure A.1 and Figure

A.2, respectively, can be put into the general framework so that there is no loss of generality.

z w

P

128 Chapter A. Robust Control Theory

Expressing the transfer functions that appear in the general framework can be done in an elegant

way using the linear fractional transformation (LFT). Assuming compatible dimensions the upper

and lower LFT are defined as:

tioned as

M11 M12

M=

M21 M22

and let be another complex matrix. Then we can define the lower LFT with

provided the inverse (I M22 )1 exists. Similarly, we can define the upper LFT with

The use of the LFT for transfer functions should be clear from the block diagram representations

in Figure A.5. The closed-loop transfer functions are given as Tz1 w1 = F
(M, ) and Tz2 w2 =

Fu (M, ), respectively.

F
(M, ) Fu (M, )

z1 w1

M

z2 M w2

Synthesis Analysis

Applying the LFT formulas to the general framework in Figure A.4 gives for the closed-loop

system

For analysis purposes of a given control system, the controller K in the general framework from

Figure A.4 can be combined with the plant P, which results in the right block diagram in Figure A.5

with M(s) = F
(P(s), K (s)). The controller synthesis results in the left block diagram in Figure

A.2. Controller Synthesis 129

A.5, which is redrawn in Figure A.6 for future reference with the controller K and G is chosen

depending on the design task:

P22 P23

nominal performance ( = 0): G =

P32 P33

P11 P13

robust stability: G=

P31 P33

P11 P12 P13

robust performance: G = P = P21 P22 P23

P31 P32 P33

Based on the background presented in this section, the synthesis of controllers will be discussed

in the next.

z w

G

y u

A.2.1 H Optimal Control

Referring to the synthesis framework in Figure A.6 the H optimal control problem becomes

min

F
(G, K )

= min

Tzw

K K

where the minimisation takes place over all internally stabilising and proper (admissible) controllers

K . Finding the optimal H controller is often numerically and theoretically complicated, which

is in contrast to the standard H2 problem, where the optimal solution is unique and can be found

solving two Riccati equations without iterations. However, in practice it is often sufficient to find a

suboptimal H controller. The control problem then becomes:

Definition A.3 (Suboptimal H Control:) Given > 0, find all admissible controllers

K (s), if there are any, such that

Tzw

< .

The synthesis of such controllers is based on the solution of two Riccati equations. The algorithm

can then be used iteratively to reduce until the minimum is reached within a given tolerance.

130 Chapter A. Robust Control Theory

A.2.2 H Loopshaping

The H loopshaping by McFarlane and Glover [MG92] combines two design techniques. Firstly,

the plant is shaped with a pre- and/or postcompensator according to guidelines that are taught in

undergraduate courses for SISO systems and that can be extended to MIMO systems. This way,

engineers can use their experience to obtain a trade-off between performance and robust stability.

Secondly, the stability radius of the shaped plant with an assumed coprime factor uncertainty is

optimised using the H methodology.

The H theory of the previous section can be used to solve the robust stabilisation problem for a

left coprime factor perturbed plant given by

P = ( M +

M )1 ( N +

N)

N ,

where M, M,

N RH and M, N are a left coprime factorisation of the plant P = M 1 N

as described in Section A.1.3. The block diagram for this kind of uncertainty is given in Figure

A.7, where the controller K is assumed to internally stabilise the nominal system. Note that the

subsequent results can be similarly derived using the right coprime factorisation by duality.

z1 z2

N

M

w

r y

K N M 1

The reason for using coprime factor uncertainty is that it includes a wide class of other un-

structured uncertainty types (e. g. additive and multiplicative uncertainty) and in particular allows

for variable numbers of right-half plane poles and zeros. The robust stability condition for the left

coprime factor perturbed plant is

! !

! K ! ! !

! ! ! !

! (I + P K )1 M 1 ! 1/ for ![ N M ]! < . (A.1)

! I !

This can be derived e. g. using the Small Gain Theorem A.2. If we use a normalised left coprime

factorisation, i. e. M M"+

N N"=

I , then the condition in (A.1) can be shown to be equivalent to

! !

! K !

! !

! (I + P K )1 I P ! 1/ := 1/b P,K (A.2)

! I !

A.2. Controller Synthesis 131

where b P,K can be thought of as a generalised stability margin provided K stabilises P (otherwise set

b P,K = 0). Note that b P,K can be related to the classical gain and phase margins for SISO systems.

Moreover, the optimal stability margin can be calculated as

# ! !2

! !

bopt := sup b P,K = 1 ! N M ! (A.3)

K H

where

H denotes the Hankel norm. The controller K that achieves bopt is the solution to the

corresponding H control problem described in Section A.2.1. Note that bopt is always smaller than

one. Moreover, since the coprime factorisation is normalised, the H optimal solution can be found

explicitly without iteration and bopt is known apriori from (A.3)

1. Scale the inputs and outputs of the plant according to the desired output decoupling and

actuator usage. This is especially important if signals are measured in different units that

differ in magnitudes. Scaling the plant achieves the appropriate relative importance of the

signals in the cost function to be minimised. A badly scaled plant is equivalent to a badly

formulated problem.

2. Shape the plant using a precompensator W1 and/or postcompensator W2 to give a desired open-

loop shape. Combine the plant and the shaping functions to the shaped plant Ps = W2 PW1 as

shown in Figure A.8. W1 and W2 have to be chosen such that Ps contains no hidden modes.

3. Calculate the optimal stability margin of the shaped plant bopt (Ps ). If bopt (Ps ) 1 the

specified loop-shape is incompatible with robust stability requirements and W1 and W2 have

to be redesigned.

4. Synthesise a stabilising controller K using the H design methodology for an bopt (Ps ).

5. Combine the resulting H controller with the shaping functions W1 and W2 to the final

controller K = W1 K W2 as shown in Figure A.8.

Ps

K W1 P W2

K

W2 K W1 P

132 Chapter A. Robust Control Theory

Note that due to the design procedure, the inclusion of K in the open-loop transfer function

can cause deterioration in the open-loop shape specified by Ps . However, it can be shown that the

parameter not only indicates the robust stability but also how well the design meets the open-loop

specifications. Generally, values of > 0.3 are satisfactory. Moreover, note that the set of plants

that are stabilised by a loopshaping controller can be characterised precisely in terms of the gap

metric (cf. A.3.2).

A.3.1 Structured Singular Value

So far, the uncertainty description has been assumed to be unstructured. However, uncertainties at

different locations in the system introduce a certain structure. Covering this with an unstructured

uncertainty may lead to conservative results since we may now not be able to create a destabilising

even if

Tzw

> . Moreover, since a performance objective in the H framework can be refor-

mulated as a robust stability problem, combining a robust stability requirement with a performance

objective also gives rise to a structured uncertainty.

In analysis or synthesis problems with structured uncertainties, the H norm, which is defined

as the largest singular value will be replaced by the structured singular value . To motivate the

definition of the latter, note that the largest singular value of a transfer matrix M pq can be

defined as

1

(M) :=

inf{ () : det(I M) = 0, q p } .

If we now restrict the to be in a set nn that describes the uncertainty structure we obtain

the following definition for the structured singular value:

1

(M) :=

min{ () : , det(I M) = 0}

Calculating the exact value is only possible for a few cases. Usually, an upper bound has to be

calculated by solving a convex optimisation problem. However, the resulting upper bound is not

always equal to . For the controller synthesis in the structured uncertainty framework, the D-K

iteration design methodology usually gives good results. However, there is no proof of convergence

for the involved iteration.

As described in the previous sections, robust control is concerned with a set of plants, i. e. a nominal

plant plus uncertainty, rather than with just a nominal plant. Therefore, it would be very helpful

A.3. Advanced Concepts 133

to have a measure of the distance between two systems, e. g. a nominal and a perturbed system.

Moreover, the measure should indicate the range of plants that a given controller and feedback

configuration can stabilise.

Due to its general nature, coprime factor uncertainty can be used to define a gap metric. In

the (left) coprime factorisation of a plant P = M 1 N , the coprime factors {N , M}

are not unique,

indeed the set of all possible coprime factors is given as {QN , Q M},

where Q is any function such

that Q, Q 1 H . To determine the distance of P to another (perturbed) plant P1 , we write P1 as

P1 = M 11 N 1 = (Q M +

M )1 (Q N +

N ).

N,

Varying the Q therefore changes M . Now, the gap between P and P1 can be defined as

Definition A.5 (Gap Metric) The gap between two plants P and P1 is defined as

$! !

! M !

g (P, P1 ) := inf ! N ! :

Q,Q 1 H

%

P1 = (Q M +

M )1 (Q N +

N ), P = M 1 N .

The gap metric has a direct relation to the loopshaping control from Section A.2.2, which is also

based on coprime factor uncertainty. A loopshaping controller that stabilises a nominal plant P with

a generalised stability margin of bP,K = will stabilise any plant P1 at a distance less than from

the nominal, i. e. P1 : g (P, P1 ) < . However, this result is conservative. It can be improved by the

so-called -gap metric, which tightens the results in the sense that always (P, P1 ) g (P, P1 ).

Using the -gap metric, it can be shown that all plants further away than ( (P, P1 ) > ) will be

destabilised by some compensator that stabilises the nominal. However, this result is still conservative

in the sense that there are infinitely many plants further away than , which will still be stabilised.

For simultaneous perturbation of plant and controller, the following theorem holds:

Theorem A.3 (Robust Stability in Gap Metric) Suppose the feedback system with the

pair (P0 , K 0 ) is stable. Then

Theorem A.3 gives a bound on the deterioration of the robust performance of a system with simul-

taneously perturbed plant and controller according to the -gaps between nominal plant/controller

and perturbed plant/controller. It also holds for the gap metric from Definition A.5, but again it is

more conservative. A more extensive treatment of the -gap metric can be found in [Vin99].

134 Chapter A. Robust Control Theory

B

Test Facilities

The Engineering Department at Cambridge University has two state-of-the-art (as of April 2003)

transient engine cells. These facilities were jointly funded by the Engineering and Physical Sciences

Research Council (EPSRC) and the Ford Motor Company, with approximatel 65 % of the funding

from the latter. This chapter is divided into the description of the hardware, the software, and the

interfacing between the two.

B.1 Hardware

B.1.1 Dynamometer

The dynamometer used in the diesel cell is a low inertia (0.46 kgm2 ) 125 kW AC electric motor

(400 Nm, 6500 rpm). The dynamometer is used to absorb energy from the engine or to motor it.

Typically, the dynamometer will be in speed control mode. If the engine is generating net positive

torque at the reference speed, the dynamometer will absorb energy from the shaft. Conversely, if net

negative torque is generated by the engine, the dynamometer will motor the engine to keep the speed

at the desired value. The low system inertia and fast control loop enable relatively fast transient

response characteristics.

B.1.2 Engine

The engine is a standard production Ford Puma 2.0-litre 16 valve common rail diesel engine (130

bhp @ 3800 rpm, 325 Nm @ 1800 rpm) which is equipped with a variable geometry turbocharger

(VGT) and exhaust gas recirculation (EGR). It is connected to the dynamometer via a pneumatically

actuated clutch, which can be used to accurately duplicate gear changes, and also to allow for the

engine to idle disconnected from the dynamometer.

B.2 Software

The engine cell is equipped with the MATLAB/Simulink /dSPACE suite of rapid controller

prototyping tools that are familiar to most control engineers. The Real-Time Workshop (also

a Mathworks product like MATLAB and Simulink) translates the Simulink diagram into C code,

which is compiled and run on a dedicated real-time processor by dSPACE. dSPACE provides various

135

136 Chapter B. Test Facilities

interfacing cards and Simulink blocks to represent them. The compiled code then runs independently

on the dSPACE hardware. Data are passed to and from the main computer where software enables

display and online control of variables in the code or the real-time processor. The dSPACE facilities

are summarised in Table B.1.

DS1005 Processor board PowerPC 750, 480 Mhz

DS2003 Analog input board 32 x 16 bit channels ADC

DS2103 Analog output board 32 x 14 bit channels DAC

DS2201 Multi I/O board 20 x 12 bit ADC, 8 x 12 bit DAC

16 DIO channels

DS4002 Digital I/O board 32 DIO channels with timing

DS4302 CAN interface 4 channels

The engine control unit (ECU) is interfaced via the CAN bus [YC01]. This solution allows access

to previously specified internal variables in order to monitor the standard engine sensors (e.g. mass

air flow (MAF), manifold absolute pressure (MAP), and EGR position), and to overwrite control

signals for the actuators, e.g. EGR and VGT actuators using dSPACE. The dynamometer speed is

also controlled via dSPACE.

Cylinder pressure Kistler 6123 100 kHz (-3 dB)

General pressure Strain gauge bridge 1 ms

Temperatures 1.5 mm tip K-type 300 ms t063 in water

Engine position Optical encoder 1 degree resolution

HC Cambustion fast FID 5 ms td , 5 ms t1090

NO Cambustion fast CLD 5 ms td , 5 ms t1090

Smoke AVL 415S steady-state device

Opacity Celesco 107 10 ms t1090

AFR UEGO sensors various

(Horiba, ETAS, NTK)

AFR, O2 , HC Horiba EXSA-1500 510 s td , 510 s t1090

NOx , CO, CO2

Fuel quantity AVL 733S 10 Hz update rate

Engine torque Surface Acoustic Wave 5 kHz (-3dB)

VGT position linear displacement

VGT speed inductive

In addition to the standard engine sensors, the engine is equipped with various temperature and

pressure sensors as well as with a turbocharger speed and inline shaft torque sensor. An overview is

given in Table B.2.

B.3. Hardware/Software Interface 137

Due to the fast response time of the pressure sensors and the MAF sensor, these measurements

are very fluctuating in the pulsating flow of the engine. In order to reduce measurement errors due to

improper sampling these signals are sampled each crank degree and then averaged over each engine

cycle.

138 Chapter B. Test Facilities

C

Engine Model Parameters

cv 0.7274 kJ/kg/K specific heat at constant volume

D 0.043 turbine blade diameter

1.3946 - specific heat ratio (cp /cv )

ic 0.73 - Intercooler efficiency

r 0.65 - EGR cooler efficiency

2

Jeng 0.2735 kgm engine inertia (crankshaft and flywheel)

2

Jt 3.0e-5 kgm turbocharger inertia

n cyl 4 - number of cylinders

pa 101.3 kPa ambient pressure

R 0.2870 kJ/kg/K gas constant (cp cv )

Ta 293 K ambient temperature

Tcool 360 K engine coolant temperature

3

Vd 2.0 dm total engine displacement volume

3

Vi 6.0 dm volume of the intake manifold

3

Vx 1.0 dm volume of the exhaust manifold

139

140 Chapter C. Engine Model Parameters

D

Signal Ranges on NEDC

Ambient conditions for these measurements: p=100 kPa, T=286 K. The data are obtained from

measurements with the engine under consideration using a test cell drivecycle simulator.

Tl -35 140 Nm load torque

Wci 17 217 kg/h mass air flow (MAF)

px 103 175 kPa exhaust manifold pressure

pt 26 kPa exhaust back pressure

pratt 1 1.68 - pressure ratio over turbine

Tx 380 750 K exhaust manifold temperature

141

142 Chapter D. Signal Ranges on NEDC

Bibliography

less Rich Diesel Combustion by Reducing Temperature, SAE, Paper 2001-01-0655,

2001.

Turbocharged Common-Rail Diesel Engine, 3rd International Conference on Control

and Diagnostics in Automotive Applications, Sestri Levante, Italy, 2001.

tems, IEEE Transactions on Control Systems Technology, Vol. 6, No. 1, pp. 21-32,

1998.

Toolbox, The MathWorks, Inc., Natick, MA, 1994.

Catalysts Systems to Reduce Emissions of Lean Running Gasoline Engines, SAE,

Paper 962045, 1996.

J. P. Warren. An Integrated SCR and Continuously Regenerating Trap System to Meet

Future NOx and PM Legislation, SAE, Paper 2000-01-0188, 2000.

[CA96] T. Chikahisa, T. Araki. In-Cylinder Control of Smoke and NOx by high turbulent Two-

Stage Combustion in Diesel Engines, SAE, Paper 962113, 1996.

ence on Control Applications, Glasgow, Scotland, 2002.

Engines for Controller Design, SAE, Paper 2001-01-1242, 2001.

Injury, J. Aerosol Sci, Vol. 29, No. 5/6, pp. 553-560, 1998.

143

144 Bibliography

New York, 1992.

to Standard H2 and H Control Problems, IEEE Transactions on Automatic Control,

Vol. AC-34, No. 8, pp. 831-847, 1989.

[FWA01] Z. Filipi, Y. Wang, D. Assanis. Effect of Variable Geometry Turbine (VGT) on Diesel

Engine and Vehicle System Transient Response, SAE, Paper 2001-01-1247, 2001.

C. K. Westbrook. Diesel Combustion: An Integrated View Combining Laser Diagnos-

tics, Chemical Kinetics, and Empirical Validation, SAE, Paper 1999-01-0509, 1999.

[GNL+95] P. Gahinet, A. Nemirovski, A. J. Laub, M. Chilali. LMI Control Toolbox, The Math-

Woirks, Inc., Natick, MA, 1995.

M. Crocker. Advanced Urea SCR Catalysts for Automotive Applications, SAE, Pa-

per 2001-01-0514, 2001.

Improved SCR Systems for Heavy Duty Applications, SAE, Paper 2000-01-0189, 2000.

[GBB+00] M. Guyon, P. Blanche, C. Bert, L. Philippe, I. Messaoudi. NOx -Trap System Develop-

ment and Characterization for Diesel Engines Emission Control, SAE, Paper 2000-01-

2910, 2000.

[GA98] L. Guzzella, A. Amstutz. Control of Diesel Engines, IEEE Control Systems Magazine,

Vol. 18, No. 5, pp. 53-71, 1998.

and Bumpless Transfer Method, Automatica, Vol. 23, No. 6, pp. 729-739, 1987.

Catalyst Concept for European Ultra-Low Emission Heavy-Duty Diesel Engines, SAE,

Paper 952652, 1995.

1988.

[Hoa01] J. Hoard. Plasma-Catalysis for Diesel Exhaust Treatment: Current State of the Art,

SAE, Paper 2001-01-0185, 2001.

[HSM+00] S. Hodjati, F. Semelle, N. Moral, C. Bert, M. Rigaud. Impact of Sulphur on the NOx Trap

Catalyst Activity-Poisoning and Regeneration Behaviour, SAE, Paper 2000-01-1874,

2000.

145

Proceedings of the IFAC World Congress, Beijing, China, 1999.

bocharged Diesel Engines, Proceedings of the American Control Conference, Philadel-

phia, PA, 1998.

Modeling of a Small Turbocharged Diesel Engine, SAE, Paper 910070, 1991.

[Joh01] T. V. Johnson. Diesel Emission Control in Review, SAE, Paper 2001-01-0184, 2001.

[Joh00] T. V. Johnson. Diesel Emission Control Last 12 Months in Review, SAE, Paper

2000-01-2817, 2000.

[KM95] M. Kao, J. J. Moskwa. Turbocharged Diesel Engine Modeling for Nonlinear Engine

Control and State Estimation, ASME Journal of Dynamic Systems, Measurement, and

Control, vol. 117, 1995.

[KLF00] M. Khair, J. Lemaire, S. Fischer. Achieving Heavy-Duty Diesel NOx /PM Levels Below

the EPA 2002 Standards An Integrated Solution, SAE, Paper 2000-01-0187, 2000.

Vol. 29, No. 5/6, pp. 575-588, 1998.

P. Herzog. NOx -Nachbehandlung fr Diesel-Pkw gelst? Entschwefelung von NOx -

Speicher-Katalysatoren, 22. Internationales Wiener Motorensymposium, Vienna, 2001.

Modelling and Control of Intake Flow in Variable Geometry Turbocharged Engines,

Proccedings of the 18th IFIP Conference on System Modelling and Optimization, De-

troit, MI, 1997.

Exhaust Gas Aftertreatment Concepts for Passenger Car Diesel Engines with Sulphur-

free Fuel, SAE, Paper 1999-01-1328, 1999.

H. P. A. Calis. Bench-Scale Demonstration of an Integrated DeSoot-DeNOx System,

SAE, Paper 2001-01-0515, 2001.

in Diesel Combustion, SAE, Paper 980184, 1998.

Thermal Effects of Exhaust Gas Recirculation on Diesel Engine Emissions Part 1:

Effect of Reducing Inlet Charge Oxygen, SAE, Paper 961165, 1996.

146 Bibliography

Thermal Effects of Exhaust Gas Recirculation on Diesel Engine Emissions Part 2:

Effects of Carbon Dioxide, SAE, Paper 961167, 1996.

Thermal Effects of Exhaust Gas Recirculation on Diesel Engine Emissions Part 3:

Effects of Water Vapour, SAE, Paper 971659, 1997.

Thermal Effects of Exhaust Gas Recirculation on Diesel Engine Emissions Part 4:

Effects of Carbon Dioxide and Water Vapour, SAE, Paper 971659, 1997.

Blended Fuel in Reducing Emissions by Varying Injection Timing or After-Treatment

Device, SAE, Paper 2001-01-0513, 2001.

of Freedom Controllers, Automatica, Vol. 29, No. 1, pp. 157-168, 1993.

Ultra Low Emission Diesel Vehicles, SAE, Paper 1999-01-0108, 1999.

bines under Pulsating Flow Conditions, SAE, Paper 2001-01-0272, 2001.

Catalyst, SAE, Paper 950374, 1995.

IEEE Transactions on Automatic Control, 37(6):759-769, 1992.

High Thermal Efficiency, and Low Noise Diesel Combustion with Oxygenated Agents

as Main Fuel, SAE, Paper 980506, 1998.

M. Sugiura, K. Kasahara. Development of New Concept Three-Way Catalyst for Au-

tomotive Lean-Burn Engines, SAE, Paper 950809, 1995.

[Moo87] J. F. Moody. Variable Geometry Turbocharging with Electronic Control, SAE, Paper

860107, 1987.

plications, SAE, Paper 1999-01-0906, 1999.

147

a Current to Vacuum Transducer and VNT Actuator, 1999 IEEE/ASME International

Conference on Advanced Intelligent Mechatronics, Atlanta, GA, 1999.

Variable Geometry Turbocharged Diesel Engine, COSY Workshop Proceedings of ECC

97, Brussels, 1997.

Ignition Engines, SAE, Paper 980784, 1998.

for Diesel Engines: An Experimental Comparison, SAE, Paper 2000-01-0266, 2000.

M. Criddle. Decentralized and Multivariable Designs for EGR-VGT Control of a Diesel

Engine, IFAC Workshop on Advances in Automotive Control, Mohican State Park, OH,

1998.

[PG97] G. Papageorgiou, K. Glover. Design and Development of a Wind Tunnel Model for High

Angle of Attack Active Control Experiments, European Control Conference, 1997.

C. W. Heath. Particulate Air Pollution as a Predictor of Mortality in a Prospective Study

of U.S. Adults, Am. J. Respir. Crit. Care Med., Vol. 151, pp. 669-674, 1995.

[RGG00] M. F. Russell, G. Greeves, N. Guerrassi. More Torque, Less Emissions and Less Noise,

SAE, Paper 2000-01-0942, 2000.

[SMB00] O. Salvat, P. Marez, G. Belot. Passenger Car Serial Application of a Particulate Filter

System on a Common Rail Direct Injection Diesel Engine, SAE, Paper 2000-01-0473,

2000.

[SA91a] J. S. Shamma, M. Athans. Gain Scheduling: Potential Hazards and Possible Remedies,

IEEE Control Systems Magazine, Vol. 12, No. 3, pp. 101-107, 1991.

[SA91b] J. S. Shamma, M. Athans. Guaranteed Properties of Gain Scheduled Control for Linear

Parameter-Varying Plants, Automatica, Vol. 27, No. 3, pp. 559-564, 1991.

ear Parameter-Varying Transformations, Journal of Guidance, Control, and Dynamics,

Vol. 16, No. 2, pp. 256-263, 1993.

[SP96] S. Skogestad, I. Postlethwaite. Multivariable Feedback Control, John Wiley & Sons,

Chichester, 1996.

148 Bibliography

in Lean NOx Catalysis for Gasoline Fueled Passenger Cars in Europe, SAE, Paper

962047, 1996.

Engine Model for Control Purposes, 3rd IFAC Workshop on Advances in Automotive

Control, Karlsruhe, 2001.

[Vin99] G. Vinnicombe. Uncertainty and Feedback: H Loop-Shaping and the -Gap Metric,

Imperial College Press, London, 1999.

[WYP+96] F. Wu, X. H. Yang, A. Packard, G. Becker. Induced L2 -Norm Control for LPV Systems

with Bounded Parameter Variation Rates, International Journal of Robust and Nonlinear

Control, Vol. 6, pp. 983-998, 1996.

[YC01] Y. Yacoub, A. Chevalier. Rapid Prototyping with the Controller Area Network (CAN),

SAE, Paper 2001-01-1224, 2001.

[You01] P. M. Young. Structured Singular Value Approach for Systems with Parametric Uncer-

tainty, International Journal of Robust and Nonlinear Control, Vol. 11, pp. 653-680,

2001.

[Zam81] G. Zames. Feedback and Optimal Sensitivity: Model Reference Transformations, Mul-

tiplicative Seminorms, and Approximate Inverses, IEEE Transactions on Automatic

Control, Vol. AC-26, pp. 301-320, 1981.

[ZD98] K. Zhou, J. C. Doyle. Essentials of Robust Control, Prentice-Hall, New Jersey, 1998.

[ZDG96] K. Zhou, J. C. Doyle, K. Glover. Robust and Optimal Control, Prentice-Hall, New

Jersey, 1996.

## Bien plus que des documents.

Découvrez tout ce que Scribd a à offrir, dont les livres et les livres audio des principaux éditeurs.

Annulez à tout moment.