Vous êtes sur la page 1sur 13

Int. J. Simulation and Process Modelling, Vol. 10, No.

2, 2015 179

Bond-graph-based controller design for the


quadruple-tank process

Matías A. Nacusse* and Sergio J. Junco


LAC, Laboratorio de Automatización y Control,
Departamento de Control,
Facultad de Ciencias Exactas, Ingeniería y Agrimensura,
Universidad Nacional de Rosario,
Ríobamba 245 Bis – S2000EKE Rosario, Argentina
Email: nacusse@fceia.unr.edu.ar
Email: sjunco@fceia.unr.edu.ar
*Corresponding author

Abstract: The quadruple-tank process has been proposed as a benchmark for multivariable
control system design. This paper addresses the design in the bond-graph domain of a robust
controller having the volumetric flows of two pumps as manipulated variables and the level of
the two lower tanks as the regulated outputs. The basic control objectives are expressed in terms
of desired closed-loop energy and power-dissipation functions and captured in the bond-graph
domain by means of a so-called target bond-graph. A basic controller design performed via
bond-graph prototyping yields a primary control law which is further robustified against
parameter uncertainties, measurement deviations and faults using the diagnostic bond-graph
concept. This results in an additional closed-loop consisting of a PI-law which is represented by a
physically meaningful bond-graph subsystem. The design methodology is first developed on a
simpler two-tank SISO-control problem and then straightforwardly extended to the multivariable
problem with the help of some causal manipulations on the four-tank bond-graph model.

Keywords: quadruple-tank system; bond-graph prototyping; nonlinear energy-based control;


robust fault-tolerant control.

Reference to this paper should be made as follows: Nacusse, M.A. and Junco, S.J. (2015)
‘Bond-graph-based controller design for the quadruple-tank process’, Int. J. Simulation and
Process Modelling, Vol. 10, No. 2, pp.179–191.

Biographical notes: Matías A. Nacusse received his degree in Electronic Engineering from the
Universidad Nacional de Rosario (UNR), Argentina, in 2007. Since April 2008 he has been a
PhD student in Electronic Engineering and Control at the Faculty of Engineering (FCEIA) of
UNR under the supervision of Prof. S. Junco and Prof. M. Romero. His work is supported by The
Argentine National Council of Scientific and Technical Research (CONICET). He was a
Teaching Assistant in two undergraduate courses, one on system dynamics and control and the
other on control of electrical drives, both at FCEIA-UNR. His main research interests are on
bond graphs, fault tolerant control and nonlinear control.

Sergio J. Junco received his Electrical Engineering degree from the Universidad Nacional de
Rosario (UNR), Argentina, in 1976. From 1976 to 1979, he worked as an Automation-Project
Engineer at Acindar, a large private steel company in Argentina. From 1979 to 1981, he was at
the Institute of Automatic Control of the University of Hannover, Germany, with a scholarship
from DAAD (the German Academic-Exchange Service). In September 1982, he joined UNR,
where he currently is a Full Professor at the recently created Department of Control. He teaches
courses on modelling and simulation of dynamical systems and control of electrical drives. He
has held several invited positions at research labs and universities in Spain and France. He is a
member of AADECA, IEEE and IFAC (individual member). His current research interests are in
theoretical and application problems in modelling, simulation, control and diagnosis of dynamic
systems.

This paper is a revised and expanded version of a paper entitled ‘Bond-graph-based controller
design of a two-input two-output four-tank system’ presented at the IMAACA 2013 Conference,
Athens Greece, 25–27 September 2013.

Copyright © 2015 Inderscience Enterprises Ltd.


180 M.A. Nacusse and S.J. Junco

1 Introduction concepts commonly found in the design of AFTC


controllers are used.
The quadruple-tank system proposed in Johansson (2000)
The active approach modifies the control law according
became a very popular benchmark allowing to test different
to the faults occurred, so that in this approach a fault
control algorithms for multivariable processes. It is a
detection and isolation (FDI) phase is mandatory before
two-input two-output or TITO nonlinear plant consisting of
making a decision on how to reconfigure the control law.
four interconnected water tanks fed by two pumps, whose
Analytical redundancy relationships (ARRs), which count
linearised model has a multivariable zero, which can be
among the many solutions used to generate residual signals
made minimum or non-minimum phase by simply changing
for FDI, have been implemented in the BG domain via the
the position of two distribution valves. In Roinila et al.
diagnostic bond-graphs (DBGs) technique presented in
(2008), a correction of the linearised model of Johansson
Samantaray et al. (2006). Using the plant inputs and plant
(2000) is presented, showing a remarkable discrepancy
measurements, residual signal are generated that depend on
when the pumps used are not identical. In Johnsen and
the model parameters and the real plant parameters.
Allgöwer (2007), an interconnection and damping
This article collects and extends the results addressed in
assignment plus passivity-based control (IDA-PBC)
Nacusse and Junco (2011, 2013), two papers presented at
algorithm for the minimum phase configuration of the
the conferences IMAACA 2011 and 2013, respectively. The
four-tank system is presented showing simulation and
first one deals with the PFTC problem on a two-tank system
experimental results. Here, the control problem is first
in the BG domain using an energy and power shaping
tackled on the two-tank configuration, where some
method (Junco, 2004). This method first expresses the
matching equations are easier to solve, and then extended to
control system specifications in terms of desired closed-loop
the four-tank system. In both cases, the level of all the tanks
energy and power dissipation functions, proceeds further
are measured. In Biswas et al. (2009), a controller design
capturing both functions in a so called target bond-graph
based on the coupling of feedback linearisation with sliding
(TBG) that represents the desired closed-loop behaviour,
mode yields an algorithm providing robust control of the
and concludes constructing the controller via bond-graph
process. In particular, the controller is designed in the
prototyping. This prototyping is such that the coupling of
difficult zone of controllability, the non-minimum phase
the resulting controller-BG and the plant-BG renders the
operating zone of the process. However, it should be noted
whole equivalent to the TBG. The basic control law
that a singularity was encountered when using this
obtained in this way is further robustified with additional
controller, when one of the four tanks is empty. Hence, the
terms derived considering a DBG of the closed-loop: the
proposed controller cannot be implemented in such case.
nominal control system represented by the TBG (originally
Limon et al. (2010) present a robust model predictive
proposed under ideal assumptions) is fed with the actual
control (MPC) algorithm for level tracking implemented in
reference signals and measured plant outputs; see Appendix
the tank system, with the plant modelled as a linear system
for a brief description of the DBG for FDI purposes. Thus,
with additive uncertainties. In Abdullah and Zribi (2012), a
the residual signal obtained from the closed-loop DBG
bibliographical review and three different control schemes
(CL-DBG) is a measure of the error between the desired and
are presented only for the minimum phase configuration:
the actual dynamics of the control system. Then, the control
gain scheduling, linear parameter varying and input-output
law is designed in order to make the residual signal vanish
feedback linearisation controllers have been compared,
in time, thus forcing the closed-loop system to behave
measuring the pressure of the four tanks.
asymptotically like the original TBG. Nacusse and Junco
This paper addresses the design in the bond-graph (BG)
(2013) extend the previous results to the four-tank problem.
domain of an energy-based nonlinear control law which
The rest of this paper is organised as follows. Section 2
only measures the pressures of the two bottom tanks. The
presents the system properties. Sections 3 and 4 revisit and
control system design objectives are to track the levels (or
extend the results and methods previously addressed in the
pressures) of the two-bottom tanks, to reject disturbances
above mentioned conference papers. Specifically, Section 3
originated in model uncertainties and measurement errors,
is devoted to the PFTC problem on the two-tank system
and to be tolerant to some interconnection faults. Fault
while Section 4 performs some BG manipulations that
tolerant control (FTC) can be classified in two main
provide a way to extend the design method employed in the
categories, passive fault tolerant control (PFTC) and active
simpler system to the multivariable four-tank process.
fault tolerant control (AFTC). Both approaches are usually
Section 5 presents some simulation results showing the
complemented in the praxis to improve the performance and
good dynamic response of the control system and, finally,
stability of the fault tolerant system (Blanke et al., 2006).
Section 6 addresses some conclusions. Sections 3 and 4 are
Refer to Zhang and Jiang (2008) for a bibliographical and
developed in detail in order to show paradigmatically
historical review on FTC. The passive approach defines a
several analysis and control system design techniques in the
unique control law to achieve the control objectives even in
bond-graph domain, as this methodological issue is a major
the presence of a fault. Generally speaking, the passive
concern of this paper.
approach ensures stability and confers robustness under
faults to the control system, but there exists a trade-off
between performance and robustness (Isermann, 2006).
Although this paper follows a PFTC-approach, some
Bond-graph-based controller design for the quadruple-tank process 181

2 System study ⎡(1 − γ1 ) u1 + u2 γ2 ⎤⎦


2

P1 = ⎣
The quadruple-tank system depicted in Figure 1 is a TITO a12
nonlinear plant consisting of four interconnected water 2
⎡(1 − γ2 ) u2 + u1γ1 ⎤⎦
tanks fed by two pumps, whose linearised model has a P2 = ⎣
multivariable zero, which can be made minimum or a 22
2
non-minimum phase by simply changing the position of ⎡(1 − γ1 ) a 2 P2 − γ1a1 P1 ⎤ γ22 (2)
two-distribution valves. In this section, the zero dynamics is P3 = ⎣ 2

explored directly in the nonlinear model on the BG domain ⎡⎣1 − ( γ1 + γ2 ) ⎤⎦ a32
using the methodology developed in Junco (2000). 2
⎡(1 − γ2 ) a1 P1 − γ2 a 2 P2 ⎤ γ12
P4 = ⎣ ⎦
The BG model of the four-tank system is depicted in
Figure 2 where the system outputs, i.e., the pressures of the 2
⎣⎡1 − ( γ1 + γ2 ) ⎦⎤ a4
2
two-bottom tanks, are indicated with arrows and denoted as
{y1, y2}.
Figure 2 BG model of the four-tank system
The state equations can be read from the BG of Figure 2
using the standard procedure, giving as state variables the
stored liquid volumes. Here, the gauge pressures at the
bottom of the tanks are chosen as state variables and their
dynamics presented in (1):

a P a P (1 − γ1 )
P1 = − 1 1 + 3 3 + u1
C1 C1 C1
a P a P (1 − γ2 )
P2 = − 2 2 + 4 4 + u2
C2 C2 C2
(1)
a P γ
P3 = − 3 3 = 2 u2
C3 C3
a P γ
P4 = − 4 4 + 1 u1
C4 C4 To analyse the zero dynamics of the four-tank system in the
BG domain requires to identify the input-output causal
where with i = 1, 2, 3, 4, Pi represents the gauge pressure at paths of minimum length as done in Figure 2 in dashed
the bottom of Tank_i; Ci are the tanks hydraulic capacities lines, then to invert the causality on these paths as done in
and ai are coefficients depending on the cross sections of the Figure 3 using the bi-causality assignment and, finally,
outlet holes of the tanks. imposing a zero signal at the outputs variables and
propagating this imposition up to the system inputs through
Figure 1 Four-tank system with measurements encircled in red the inverted causal paths. For more details of this procedure
(see online version for colours)
see Junco (2000). Applying it in Figure 3 yields the
necessary inputs to maintain y1 = 0 and y2 = 0 according to
(3). A new BG model can be constructed for the constrained
inputs in order to illustrate the relationships between the
remaining variables, i.e., the zero dynamics, as done in
Figure 4. This figure, which is just the first step towards
the zero-dynamics bond-graph, shows the C elements
representing the two upper tanks in derivative causality in
order to put in evidence an algebraic loop closed through
the MSf’s which is convenient for further manipulation of
the BG. The second step, the elimination of the algebraic
loops, yields Figure 5, which shows the appearance of new
R elements. Finally, in order to get a better insight into the
whole dynamics, this bond-graph is converted into an
equivalent vector BG with the following manipulations: the
single-port capacitors, resistors and modulated flow sources
are respectively grouped into the C-, R- and MSf-fields
depicted in Figure 6. The constitutive relationships of these
fields are detailed in (4).
The equilibrium points of the four-tank system must satisfy
the constraints (2). This restriction on the tank state
variables limits the regulation problem up to two tank
pressures (or levels).
182 M.A. Nacusse and S.J. Junco

Figure 5 Reduced BG model for zero dynamics study, step 2


a3 P3
u1 = −
(1 − γ1 )
(3)
a P
u2 = − 4 4
(1 − γ2 )
⎡ γ1 ⎤
⎡C4 0⎤ ⎢ 0 1 − γ1 ⎥
C=⎢ ,Γ=⎢ ⎥
⎣0 C3 ⎥⎦ ⎢ γ2 ⎥
⎢1 − γ 0 ⎥
⎣ 2 ⎦
(4) A similar analysis yielding the same results can be
⎡ 1 − ( γ1 + γ2 ) ⎤
⎢ a4 P4 ⎥ performed on an exact incremental BG around the generic
(1 − γ1 )(1 − γ2 ) equilibrium point specified in (2). Alternatively, this result
fR = ⎢ ⎥
⎢ 1 − ( γ1 + γ2 ) ⎥ can be obtained considering a linearised model around this
⎢ a3 P3 ⎥ equilibrium point.
⎣⎢ (1 − γ1 )(1 − γ2 ) ⎦⎥
The MSf-field of Figure 6 acts as a state-dependent Figure 6 Equivalent BG model for zero dynamics study
perturbation on the remaining RC-subsystem. If 1 – (γ1 + γ2)
> 0 then the R-field is strictly dissipative and the
RC-subsystem can be shown to be exponentially stable. As
the MSf-field injects a vanishing perturbation, the resultant
whole dynamics, i.e., the zero dynamics is stable. When
1 – (γ1 + γ2) < 0 the zero dynamics is unstable. When
1 – (γ1 + γ2) = 0, then fR = 0 and the system neither generates
nor dissipates the energy stored in the C field. In the
linearised system, this situation coincides with the zeros
placed at the origin of the complex plane.

Figure 3 IO-causality inverted BG with (zero) outputs as inputs 3 Background and previous results
This section briefly summarises the main ideas on
performing energy shaping and damping assignment
directly in the BG domain through BG prototyping and
recalls their application to solve a control problem on a
two-tank system as presented in Nacusse and Junco (2011).
This result will be reinterpreted, by performing a causal
manipulation, as a prelude to the development of the main
result in this paper, the design of a controller for the
quadruple-tank benchmark process.

3.1 Energy-based control in the BG domain


The power and energy shaping control technique defines the
control problem as a stabilisation one, choosing desired
closed-loop energy- and power-dissipation functions, and
Figure 4 Reduced BG model for zero dynamics study, step 1 obtaining the control law through equations that match the
control open-loop energy function [a kind of control
Lyapunov function, see Sontag (1998)] and the desired
closed-loop functions. In the BG domain, the closed-loop
specifications are expressed by a so-called TBG
representing the desired closed-loop behaviour. In order to
obtain the control law, the controlled sources – which
provide the manipulated variables in the BG model of the
plant – are prototyped (meaning that their behaviour is
expressed by BG components) in such a way that the global
BG resulting from their power-interconnection with the rest
of the plant BG – which is called a virtual bond-graph
(VBG) – matches the TBG. The control law is obtained
from the VBG by simply reading the outputs of the
Bond-graph-based controller design for the quadruple-tank process 183

prototyped sources with the help of the causal assignment in Figure 8 TBG for pressure control of lower tank
the VBG. This method is exemplarily performed below on a
two-tank system. For more details refer to Junco (2004).

3.2 Control via BG-prototyping: the 2-tank system


As shown in Figure 7, the tanks are located one above the
1
other, the upper tank discharging into the lower tank. Both V ( P) = C1 Pe2
tanks are fed with a unique input flow splitted between them 2
(6)
through a distribution valve whose parameter γ ∈ [0, 1] 1 2
V ( P) = − Pe
determines how the input flow is distributed to the tanks as RH
indicated by the BG in Figure 7.
1
The control objectives imposed on Tank_1 are: Pe = − Pe (7)
C1 RH
• regulation of constant reference levels
• rejection of constant disturbances The regulation error Pe = P1 − P1ref is the state variable of
the TBG and P1ref is the Tank_1 reference pressure.
• robustness regarding parametric uncertainties and
faults. To enforce the desired closed-loop dynamics specified
by the TGB, the VBG of Figure 9 is constructed. It shows
Figure 7 Two-tank system and its BG model (see online version how to proceed in order to obtain the control law. The left
for colours) half of the figure is obtained prototyping the controlled
power source MSf in such a way that access is gained to the
chosen output, the pressure P1, and an overall equivalent
behaviour to the TBG is achieved. The first objective is
achieved via the exact compensation of the Tank_2 pressure
on the pump over the distribution valve and of the discharge
of Tank_2 on Tank_1. The second objective is reached first
adding the virtual elements with negative ‘gains’ that cancel
the own dynamics of Tank_1 and later building the
incremental dynamics around the reference pressure P1ref
for Tank_1 via the insertion of the virtual elements C: C1
and MSe (shapes the closed-loop energy) and R: RH
(assigns damping).

Figure 9 Virtual and plant BG (see online version for colours)


Note: Measured plant outputs encircled in red.
The state equations can be read from the BG of Figure 7
using the standard procedure, giving as state variables the
stored liquid volumes. Here, the gauge pressures at the
bottom of the tanks are chosen as state variables [instead of
the liquid levels used in Nacusse and Junco (2011)] whose
dynamics is presented in (5):

a P a P (1 − γ )
P1 = − 1 1 + 2 2 + u
C1 C1 C1
(5)
a P γ
P2 = − 2 2 + u
C2 C2
where with i = 1, 2, Pi represents the gauge pressure at the
bottom of Tank_i; Ci are the tanks hydraulic capacities and Using the standard causality reading procedure the control
ai are coefficients depending on the cross sections of the law (8) can be read directly from the VBG:
outlet holes of the tanks.
⎛ 1 ⎞⎡
The proposed TBG for the closed-loop system is shown u=⎜ ⎟ ⎢ a1 P1 − a2 P2 −
1
( P1 − P1ref ) + C1 P1ref ⎥⎤ (8)
in Figure 8 where the desired stored energy (V) and power ⎝1− γ ⎠ ⎣ RH ⎦
dissipation (V ) are expressed in terms of the regulation
Assuming exact model knowledge and perfect
error state variable (Pe) in (6) and (7).
measurements, this control law yields a closed-loop
184 M.A. Nacusse and S.J. Junco

behaviour equivalent to the TBG of Figure 5, i.e., the Furthermore, as the control objectives are placed only on
closed-loop dynamics satisfies (7). Tank_1, for control system design purposes the equivalent
plant model shown in Figure 12 can be considered, where
Remark: the rated control law (8) performs a partial energy
the effect of Tank_2 enters as a disturbance.
shaping and damping assignment, since only the dynamics
of Tank_1 is captured in the TBG. As no objectives are Figure 12 Equivalent BG for control system design
imposed on Tank_2 and its dynamics is hidden in
closed-loop, its stability must be analysed after the
controller has been designed, property that can be easily
verified in this case.

3.3 Causal manipulation and simplified control law


Another manner to obtain the control law (8) is performing
a causal manipulation on the BG depicted in Figure 7. In
this figure, the flow Q21 from Tank_2 into Tank_1 is
computed by the R-element representing the discharge
orifice of the former. Putting in derivative causality the Following similar steps as those detailed in Section 3.2, the
upper C-element (models the potential energy stored in the VBG of Figure 13 can be constructed to enforce the desired
upper tank) yields the BG of Figure 10, where the closed-loop dynamics specified by the TGB.
discharging flow is computed as the sum of the incoming
flow from the source and the flow of the upper C element, Figure 13 Virtual and plant BG (see online version for colours)
i.e., Q21 = γu – Qc2. This new computation justifies the
equivalent BG of Figure 11.

Figure 10 BG model of the two tanks system with the upper C


in derivative causality (see online version for colours)

The control law (9) can be read directly from the VBG
using the standard causality reading procedure:
1
u = a1 P1 − ( P1 − P1ref ) + C1 P1ref + C2 P2 (9)
RH
Remark: Equations (8) and (9) are fully equivalent and both
show the need to measure not only the pressure of Tank_1
but also that of Tank_2 to implement the control law u.
However, equation (9) – which results from the above
manipulation performed on the BG – has the advantage of
Figure 11 Equivalent plant model after manipulation of the showing that its last term can be eliminated from the control
original BG (see online version for colours) law as it is an evanescent perturbation, i.e., it vanishes in
steady state, and as such it does not affect the equilibrium
point. Moreover, as the autonomous system shown in
Figure 12 is LTI, then the non-autonomous system with
evanescent perturbation is exponentially stable.
Doing so yields the control law (10) which can be
implemented with the sole measurement of P1. However, it
must be realised that this simplification amounts to
modifying the TBG as indicated in Figure 14.
1
u = a1 P1 − ( P1 − P1ref ) + C1 P1ref (10)
RH
Bond-graph-based controller design for the quadruple-tank process 185

Figure 14 New TBG without measuring x2 and the ideally expected closed-loop dynamics, i.e., when
res = 0, Pe responds as previously defined in the TBG of
Figure 8. This suggests that the control objectives could be
reached extending the previously computed control law to a
new one u = u ( P1 , P1ref , res ) incorporating the residual
signal in such a way that res tends to zero with growing
time.

Perturbed closed-loop dynamics 1 res


Pe = − Pe + (13)
C1 RH C1
Because of parameter dispersion, faults, modelling errors,
sensor limited precision, noise, etc., neither the model nor The residual expression (14) obtained reading the CL-DBG
the measurements are exact. To deal with this it is clearly shows that choosing u as in (8) (rated control law)
convenient to think the control input as composed by two yields res = 0, in absence of faults and modelling errors.
terms as in (11), where ur is the ‘rated’ part of u, i.e., the
1
control input part that performs the power and energy res = −a1 P1 + ( P1 − P1ref ) − C1 P1ref − C2 P2 + u (14)
shaping under ideal plant and measurement conditions, and RH
δu is the unknown controller part due to modelling errors,
To improve the control system robustness, the extra term u4
parametric dispersion, faults, etc. The perturbed target
shown in (15) is added to the expression (8) for u.
bond-graph (PTBG) of Figure 15 reflects this situation.
1
u = ur + δu (11) u = a1 P1 − ( P1 − P1ref ) + C1 P1ref + C2 P2 + u4 (15)
RH
Figure 15 PTBG

Choosing u4 = − K res yields the residual dynamics (16):

 + Kres = δu
res (16)
Thus, with constant δu, res goes asymptotically to zero with
time constant 1/K. As already anticipated, this forces Pe to
Under this situation the closed-loop dynamics no longer approach asymptotically the desired error dynamics defined
satisfies (7) but (12). in the TBG of Figure 8.

1 1 Remark: the assumption of constant δu is based in the fact


Pe = − Pe + δu (12) that the discrepancies, caused by parameter uncertainties,
C1 RH C1 faults, modelling errors, etc., between the model and the real
plant are constant. In rigour, δu is state dependent, i.e.,
3.4 Robustifying the control law δu = g δu ( P ), and being constant at the equilibrium point
The CL-DBG is defined injecting the pressure tracking error means that δu = 0 so that this term in (16) can be
(as measured on the real control system) into the TBG considered as a vanishing perturbation.
through modulated sources and collecting a residual signal
as shown in Figure 16. This residual measures the error Representing u4 in terms of ( P1 − P1ref ) yields (17),
between the PTBG and the TBG and will be used to expression showing that, in this case, the residual signal
generate an additional control action forcing this error to defined in the CL-DBG has a PI structure. Note however
vanish in steady-state (as well as the residual itself). that this does not necessarily generalise, since the resulting
structure depends on the TBG.
Figure 16 Proposed CL-DBG (see online version for colours)
K
u4 = − KC1 ( P1 − P1ref ) − ∫(P − P )
1 1
ref
(17)
RH

3.5 BG interpretation of the robustifying term


The results of the previous subsection are revisited here in
order to provide a BG implementation of the additional term
Note: Measurements to be fed encircled in red. u4 of the control law given in (17).
The CL-DBG yields the new error dynamics in (13) which The TBG defines the closed-loop dynamics of the
is driven by the residual signal PQR. The residual signal, system. To robustify its behaviour it is necessary to inject an
which is the power co-variable of the error injected into the additional control action. In order to analyse this problem in
CL-DBG, is a measure of the difference between the actual the BG domain a concise word BG version of that one in
Figure 13 rated controller is presented in Figure 17.
186 M.A. Nacusse and S.J. Junco

Departing from Figure 17, Figure 18 shows the word BG of Figure 20 Power coupling between the PTBG and the mDBG
a power interconnection proposed as a means to provide the
additional control action. There, the word BG block
named mDBG must be capable of rejecting all of the
above-mentioned disturbances.

Figure 17 Power coupling between plant BG and VBG to obtain


the TBG

Figure 21 shows the resulting closed-loop BG, where the


power interconnection among the plant BG, the VBG and
the mDBG is highlighted in dotted squares. Notice that the
redundancy in the MSe which injects the − P1ref can be
Figure 18 Proposed power interconnection eliminated by shifting the mDBG into the VBG as it is
shown in Figure 22.

Figure 21 ResultingCL-BG (see online version for colours)

The residual signal is generated through a CL-DBG as


indicated in Figure 19 (cf. Figure 16).

Figure 19 Signal coupling between PTBG and TBG through


closed-loop DBG

Figure 22 CL-BG with mDBG coupled into the VBG


(see online version for colours)

The additional control action u4 (a volumetric flow) given in


(17) in terms of the tracking error (a pressure) can be
interpreted as produced by an effort-sharing set constituted 3.6 Stability analysis
by a hydraulic resistance and an inertance, i.e., a R-I set. Figure 23 shows the complete CL-BG model of the two
Thus, u4 can be thought of as minus the sum of QR ≔ KC1Pe tank system employed to analyse the stability of the


and QI := K / RH Pe , the volumetric flows of the R and the equilibrium point. The equilibrium point (18) can be
calculated from the BG model by simply making zero the
I elements, respectively. Figure 20 shows the resulting incoming power variables of the storage elements in integral
closed-loop BG with power coupling between the PTBG causality and reading through the causal paths and the
and the mDBG, where R1 = 1/KC1 and I1 = RH⁄K. constitutive relationships of the elements. The reader must
refer to Breedveld (1984) or Junco (1993) for a detailed
Bond-graph-based controller design for the quadruple-tank process 187

description of an algorithm to obtain the equilibrium point The signals entering the upper tank and the
in a BG domain. PTBG-mDBG submodels through the MSf’s in the BG of
2
Figure 24 can be interpreted as vanishing perturbations.
⎛a γ⎞ Hence, the stability analysis can be first performed
P2 = ⎜ 1 ⎟ P1ref
⎝ a2 ⎠ disregarding the MSf, see Figure 25, and then extended to
Pe = 0 (18) the perturbed system since the origin of the incremental CL
model is asymptotically stable and the disturbances, i.e.,
QI = δu Pe / RH + C1 P1ref and C2 P2 , are bounded. The stability of
Instead of the BG of Figure 23, with equilibrium point at the autonomous system is easily verified by following the
(18), an equivalent incremental BG model can be propositions stated in Junco (2001), as the total energy in
constructed with equilibrium point at zero, where the the storages is a positive definite function of the state
systems relationships hold but for all the variables referred (energy) variables, thus being a candidate Lyapunov
to their equilibrium values. This incremental BG model is function, and, moreover, each R element is strictly
shown in Figure 24, where the incremental variables are dissipative, has it causality imposed by only one storage
defined as: ΔP2 = P2 − P2 , ΔPe = Pe − Pe and ΔQI = QI − QI . element in integral causality, and all storages impose
causality on a R-element, what means that the orbital
Figure 23 Complete CL-BG model for the two-tank system derivative of the total stored energy is a negative definite
function of the states.

Figure 25 Autonomous incremental CL-BG model for the


two-tank system

4 Application to the four-tank benchmark


Figure 24 Incremental CL-BG model for the two-tank system problem
In this section, the control law for the four-tank system is
designed. The control objectives are the same as that
proposed for the two-tank system, in this case imposed on
both of the bottom tanks.

4.1 Causal manipulation and controller design


As the first step developing the multivariable control law, a
causal manipulation is performed on the original BG of
Figure 2 as shown in Figure 26. As in the two-tank case,
again the discharge flows of the upper tanks are expressed
not as computed by their associated R-elements but as the
difference between the flows of the sources minus their
net input flows, i.e., as γ1u1 − C3 P3 and γ2 u2 − C4 P4 ,
respectively. This manipulation permits to construct the new
BG given in Figure 27. This BG exhibits two internal
variables u1* and u2* , which in the sequel are going to be
188 M.A. Nacusse and S.J. Junco

treated as virtual control inputs. Seen from the bonds 1 ⎪⎧ ⎡ P1 − P1ref


associated to these two auxiliary flow variables the u1 = − ⎨(1 − γ2 ) ⎢ a1 P1 − + C1 P1ref
γ1 + γ2 − 1 ⎩⎪ ⎣ RH1
quadruple-tank problem appears as two decoupled two-tank
problems. Hence, the multivariable problem is strongly K1 ⎤
− K1C1 ( P1 − P1ref ) − ∫ ( P − P )⎥⎦
1 1
ref
simplified, as the previously developed procedure can be RH1
first applied to each sub-problem and then the overall (21)
⎡ P2 − P2ref
control law be recovered using the causal relationships −γ2 ⎢ a2 P2 − + C2 P2ref − K 2C2 ( P2 − P2ref )
relating the auxiliary variables to the control inputs ⎣ RH 2
provided by the power-conserving structure relaying them. K2 ⎤⎫
∫ ( P − P )⎥⎦ ⎬⎭
ref
The real and the virtual control inputs are related − 2 2
RH 2
through a transformation matrix which is given in (19).
Notice that when γ1 + γ2 = 1 the transformation matrix is
singular. In this condition of the distribution valves, the 1 ⎧⎪ ⎡ P2 − P2ref
u2 = − ⎨(1 − γ1 ) ⎢ a2 P2 − + C2 P2ref
multivariable zeros (of the linearised system) are placed at γ1 + γ2 − 1 ⎪⎩ ⎣ RH 2
the origin of the complex plane (Johansson 2000).
K2 ⎤
− K 2 C2 ( P2 − P2ref ) − ∫ ( P − P )⎥⎦
2
ref
⎡(1 − γ1 )
2
⎡u1* ⎤ γ2 ⎤ ⎡ u1 ⎤ RH 2
⎢ *⎥ = ⎢ (19) (22)
⎣u2 ⎦ ⎣ γ1 (1 − γ2 ) ⎥⎦ ⎢⎣u2 ⎥⎦ ⎡ P1 − P1ref
−γ1 ⎢ a1 P1 − + C1 P1ref − K1C1 ( P1 − P1ref )
The virtual control laws, obtained following the same ⎣ RH 1

procedure as in the two-tank example ignoring the K1 ⎤⎫


∫ ( P − P )⎥⎦ ⎬⎭
ref
− 1
vanishing term, are given in (20). RH1
1

1
u1* = a1 P1 − ( P1 − P1ref ) + C1 P1ref Using (21) and (22) the stability of the hidden closed-loop
RH1 dynamics of Tank_3 and Tank_4 can be verified. Figure 27
(20)
1
u2* = a2 P2 − ( P2 − P2ref ) + C2 P2ref shows that the interconnection structure relating the bonds
associated to the actual control inputs u1,2 and the virtual
RH 2
*
control inputs u1,2 is composed of TFs, 0- and 1-junctions.
The actual control inputs u1,2 are simply obtained
As such, this interconnection structure is power conserving.
pre-multiplying (20) by the inverse of the matrix in (19).
This implies that the stability results obtained considering
Notice that the actual control law is not defined for the
the virtual inputs, i.e., for the two-tank problem, are
distribution valves configuration satisfying γ1 + γ2 = 1.
immediately valid for the actual inputs and thus can be
Figure 26 BG of the four tanks system with upper tanks in directly extended to the four-tank problem.
derivative causality (see online version for colours)
Figure 27 BG model after manipulation for controller design
(see online version for colours)

5 Simulation results
4.2 Robustifying the control law
The parameters used in the simulations, shown in Table 1,
The virtual control laws (20) are robustifyied, as in the were obtained from (Johansson 2000), where Ai are the
two-tank example, via power coupling of the mDBG or by cross section areas of the tanks, related to the tanks
simply adding a term like (17) to each equation of (20). This hydraulic capacities by the relation Ci = Ai/ρg, where ρ is
yields the real control laws given in (21) and (22). the liquid (water) density and g is the gravitational
acceleration. The parameters of the mDBG are
RH1 = RH 2 = 10 and K1 = K2 = 0.01.
Bond-graph-based controller design for the quadruple-tank process 189

Table 1 Simulation parameters In simulation Scenario_2, a non-minimum phase


configuration is tested, where the valves position are placed
Parameter Value
at: γ1r = 0.7, γ2r = 0.7. Again, two sequential faults occurred,
A1, A3 28 cm2 the first, at time T = 6,000 s where the value of the position
A2, A4 32 cm2 valve changes to γ2 = 1 which physically implies that
pump_2 injects all its flow to Tank_3; and the second, at
a1, a3 0.071 cm 2 cm3 / gr
time T = 15,000 s, where the value of the cross section area
a2, a4 0.057 cm 2 cm3 / gr of outlet hole of Tank_1 (a1) is increased by 50%.

Figure 29 Residual signals and control inputs (in cm3/s) for


Using the control laws (21) and (22) with noisy
sequential faults in simulation Scenario_1 (see online
measurements of the bottom tanks pressure (normal version for colours)
distribution and amplitude n = 2 cm), the simulation
scenarios involve abrupt faults in the system. To show the
robustness of the control laws, the parameters of Tank_1
and Tank_2 are varied in –10% and +10% respectively. For
illustration purposes, the simulations show tanks levels high
instead of tanks pressures.
The first simulation scenario, Scenario_1, consists of a
minimum phase configuration with valves positions in
γ1r = 0.3 and γ2r = 0.2. These are rated values used to
parameterise the control laws. In this scenario, two
sequential faults occurred: at time T = 2,000 s the value of
the cross section area of outlet hole of Tank_2 is increased
by 50% (forcing the same increment in a2); and at time
T = 5,000 s the value of the valve position changes to
γ1 = 0.6.

Figure 28 Tank levels responses (in cm) for sequential faults in


Scenario_1 (see online version for colours)

Figure 30 Tank levels responses (in cm) for sequential faults in


Scenario_2 (see online version for colours)

The simulation responses for Scenario_1 are shown in


Figure 28 and Figure 29, which show that despite the faults
occurrence the control system behaves as expected and the
bottom tanks levels follow their references. For both cases,
the control inputs, u1 and u2, force to zero the residual
signals in order to reject the faults.
190 M.A. Nacusse and S.J. Junco

In this scenario, the bottom tank levels follow their Rosario) and ANPCyT for their financial support through
references rejecting the disturbances originated by the faults the research projects PID-UNR Nr. 19I386 and PICT 2008
occurrence as it is shown in Figure 30. Figure 31 shows the Nr. 650, respectively.
associated residual signals and the control inputs. Here
again, the control inputs force the residual signals to remain
at zero to reject the faults. References
Abdullah, A. and Zribi, M. (2012) ‘Control schemes for a
Figure 31 Residual signals and control inputs (in cm3/s) for
quadruple tank process’, International Journal of Computers
sequential faults in simulation Scenario_2 (see online
Communications and Control, Vol. 7, No. 4, pp.594–604,
version for colours)
ISSN: 1841-9836.
Biswas, P.P., Srivastava, R., Ray, S. and Samanta, A.N. (2009)
‘Sliding mode control of quadruple tank process’,
Mechatronics, June, Vol. 19, No. 4, pp.548–561,
ISSN: 0957-4158, DOI 10.1016/j.mechatronics.2009.01.001.
Blanke, M., Kinnaert, M., Lunze, J. and Staroswiecki, M. (2006)
Diagnosis and Fault-Tolerant Control, 2nd ed.,
Springer-Verlag Berlin Heidelberg.
Breedveld, P.C. (1984) ‘A bond graph algorithm to determine the
equilibrium state of a system’, Journal of the Franklin
Institute, Vol. 318, No. 2,. pp.71–75, ISSN 0016-0032.
Isermann, R. (2006) Fault-Diagnosis Systems, Springer-Verlag
Berlin Heidelberg.
Johansson, K.H. (2000) ‘The quadruple-tank process: a
multivariable laboratory process with an adjustable zero’,
Control Systems Technology, IEEE Transactions on, May,
Vol. 8, No. 3, pp.456, 465, doi: 10.1109/87.845876.
Johnsen, J. and Allgöwer, F. (2007) ‘Interconnection and damping
assignment passivity-based control of a four-tank system’,
Lecture Notes in Control and Information Sciences.
Lagrangian and Hamiltonian Methods for Nonlinear Control
2006, Springer Berlin Heidelberg Publisher, Vol. 366,
pp.111–122.
Junco, S. (1993) ‘Stability analysis and stabilizing control
synthesis via Lyapunov’s second method directly on bond
graphs on nonlinear systems’, Proceedings of IECON’93,
Maui, HII, 17–20 November, pp.2065–2069.
Junco, S. (2000) ‘Linéarisation exacte entrée-sortie et stabilité de
6 Conclusions la dynamique des zéros directement sur bon graph des
systèmes non linéaires’, Actes de CIFA 2000, Première
This work addressed the design of a robust controller for a conference Internationale Francophone d’automatique,
multivariable four-tank system in three stages. A (partial) 4–7 juillet, Lille, France.
energy shaping and damping assignment control system Junco, S. (2001) ‘Lyapunov second method and feedback
design technique in the bond-graph domain was first applied stabilization directly on bond graphs’, Proc. ICBGM’2001,
SCS-Simulation Series, Vol. 33, No. 1, pp.137–142.
to obtain an almost-exact I/O feedback linearising controller
for a simpler two-tank problem. The controller is just Junco, S. (2004) ‘Virtual prototyping of bond graphs models for
controller synthesis through energy and power shaping’,
almost-exact because a feedback term was ignored in the Conference on Integrated Modeling and Analysis in Applied
(otherwise exact) control law in order to spare a Control and Automation (IMAACA 2004).
measurement. The second stage proceeded to robustify the Limon, D., Alvarado, I., Alamo, T. and Camacho, E.F. (2010)
previous controller to which aim a closed-loop DBG was ‘Robust tube-based MPC for tracking of constrained linear
introduced. Finally, a causal manipulation was performed systems with additive disturbances’, Journal of Process
on the BG of the quadruple-tank that permitted handling the Control, Vol. 20, No. 3, pp.248–260.
associated multivariable problem as two monovariable Nacusse, M. and Junco, S. (2011) ‘Passive fault tolerant control: a
decoupled problems, each for a simple two-tank system. bond graph approach’, Conference on Integrated Modeling
Simulation results demonstrate the good regulation and Analysis in Applied Control and Automation (IMAACA
response and the fault tolerance of the control system. 2011), Rome, Italy, 12–14 September.
Nacusse, M. and Junco, S. (2013) ‘Bond-graph based controller
design of a two-input two-output four-tank system’,
Conference on Integrated Modeling and Analysis in Applied
Acknowledgements Control and Automation (IMAACA 2013), Athens Greece,
The authors wish to thank SeCyT-UNR (the Secretary for 25–27 September.
Science and Technology of the National University of
Bond-graph-based controller design for the quadruple-tank process 191

Roinila, T., Jaatinen, A. and Vilkko, M.K. (2008) ‘Corrected As can be noted in (23), the residuals depend on system
mathematical model of quadruple tank process’, Proceedings parameters. If the model represents perfectly the controlled
of the 17th IFAC World Congress, South Korea, Vol. 17,
system, then the residual signals are zero. The differential
Part 1.
causality is an advantage in FDI, because no knowledge of
Samantaray, A., Medjaherb, K., Ould Bouamama, B.,
the initial states is necessary to evaluate the residuals.
Staroswieckic, M. and Dauphin-Tanguy, G. (2006)
‘Diagnostic bond graphs for online fault detection and
isolation’, Simulation Modelling Practice and Theory, April, Figure 32 DBG of the two tank system (see online version
Vol. 14, No. 3, pp.237–262. for colours)
Sontag, E. (1998) Mathematical Control Theory: Deterministic
Finite Dimensional Systems, 2nd ed., Springer, New York.
Zhang, Y. and Jiang, J. (2008) ‘Bibliographical review on
reconfigurable fault-tolerant control systems’, Annual
Reviews in Control, December, Vol. 32, No. 2, pp.229–252.

Appendix
Diagnostic bond-graph
The DBG was first presented by Samantaray et al. (2006)
for numerical evaluation of ARRs. The ARRs are calculated
to perform FDI in an AFTC frame.
Basically, the DBG is obtained from a BG model of the
plant injecting the plant measurements and inputs through
modulated sources. The residual signal is obtained by
Note: Plant measurements to be fed into the DBG
measuring the power co-variables of the modulated sources, encircled in red.
see Figure 32.
Reading directly from the BG the residuals are:

res1 = C1 P1 + a1 P1 + a2 P2 − (1 − γ)u


(23)
res 2 = C2 P2 + a2 P2 − γu