Académique Documents
Professionnel Documents
Culture Documents
Abstract: Gasoline Direct Injection (GDI) spark ignition engines equipped with the Common
Rail (CR) system strongly improve engine performance in terms of fuel consumption and
pollutant emission reduction. As a drawback the fuel pressure in the rail has to be kept as
constant as possible to the demanded pressure working set-points in order to achieve the
advantages promised by this technology. In this work a Model Reference Adaptive Control
(MRAC) algorithm based on the Minimal Control Synthesis (MCS) strategy is proposed to
reduce the residual pressure in the rail. Numerical results based on a CR mean value model,
previously proposed in the literature and experimentally validated, show that a very satisfactory
attenuation of the pressure ripple as well as pressure tracking are attained in different working
conditions. A quantitative comparison with a classical gain scheduling model-based control
approach confirms furthermore the effectiveness of the proposed adaptive control strategy.
Keywords: Adaptive control, Automotive control, Disturbance rejection, Common Rail, GDI
engine.
4174
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
0
1 0 ... 0
Z t
B =(0 0 ··· 0 b) ,
T
(3) KN (t) = γ |ye (τ )| dτ, γ > 0, (16)
0
T
d(t, x) = ( 0 0 · · · 0 f (t, x) ) . (4) being C the output matrix of the plant and ye the output
with f (t, x) being a time and state dependent nonlinear error as in (9). Notice that the signum of the adaptation
function which models the disturbance acting on the weights α and β in (8a), (8c), (15) and γ in (16) has to be
linear part of the plant model and unmodelled nonlinear the same of the term b in (3). Furthermore, typically all the
dynamics. Notice that for the design of the NEMCSI adaptive gains are initialized to zero, i.e. K(0) = K0 = 0,
strategy only the signum of b in (3) has to be known. KR (0) = KR0 = 0, KI (0) = KI0 = 0 and KN (0) = KN0 =
0.
The main aim of NEMCSI control is for the plant states,
x(t), to track asymptotically the states, xm (t), of a given In di Bernardo et al. (2010) it has been proven via
asymptotically stable reference model of the form a passivity based analysis that, as the classical MCS
algorithm, the control action (6) guarantees asymptotic
ẋm (t) = Am xm (t) + Bm r(t), (5) stability of the error dynamics (i.e. xe → 0 when t → ∞)
n
with xm (t) ∈ R , r(t) ∈ R being some desired reference in absence of the d(t, x)-term in (1), and robustness to
signal and Am (Hurwitz matrix) and Bm given in the same unmodelled plant dynamics and disturbance of the form
canonical form as that of the plant. (4) as well as robustness to plant parameter variations.
Moreover the two additional terms, i.e. uN (t) and uI (t)
The control action provided by the NEMCSI algorithm is in (6) have shown to increase tracking and disturbance
given as follows: rejection performance.
u(t) = uM CS (t) + uI (t) + uN (t), (6) We note that, when the reference model in (5) is chosen
where uM CS (t) is the classical MCS input which consists as the a nominal linear model of the plant controlled via
of a feedforward and a feedback action with time-varying a classical LQ (Linear Quadratic) strategy Anderson and
adaptive gains defined as Moore (1971), the control action (6) is termed as LQ-
NEMCSI (see di Bernardo et al. (2010) for further details).
uM CS (t) = K(t)x(t) + KR (t)r(t), (7)
with
4. NEMCSI CONTROL OF THE COMMON RAIL
Z t
K(t) = α ye (τ )xT (τ )dτ + βye (t)xT (t), (8a) The first step for the design of the NEMCSI adaptive
0
controller is to verify if the plant dynamics match or not
K(0) = K0 , K ∈ Rn , (8b) the structure (1). In this work we have chosen from the
Z t
literature the CR model proposed in di Gaeta et al. (2009)
KR (t) = α ye (τ )r(τ )dτ + βye (t)r(t), (8c) that describes the pressure, p(t) (bar), in the CR as a
0
function of the pump speed and electro-valve current, i
KR (0) = KR0 , KR ∈ R (8d) (A), and it can be decomposed into two terms, namely, the
and α and β being scalar adaptation weights. The output mean pressure term, p̄(t) (bar) and the residual pressure,
error ye is computed as η(t) (bar), describing the ripple around the mean value.
The CR dynamical system is then given by
ye (t) = Ce xe (t), (9)
di R Vb aδ(t) + b
where =− i+ (17a)
dt L L 100
p̄(t) = c(N )i + d(N ), (17b)
xe (t) = xm (t) − x(t), (10a) p(t) = p̄(t) + η(t; p̄, N, Tinj ), (17c)
Ce = [ 0 . . . 0 1 ] P , (10b) where: δ (%) is the duty-cycle expressed in percentage
and P is the solution of the Lyapunov equation terms of the PWM signal used to actuate the electro-valve
((8) in Fig. 1); a and b are parameters of the electro-valve
P Am + ATm P = −M, M > 0. (11) actuation circuit; L (H) and R (Ω) are the electro-valve
inductance and the electric resistance of coil respectively,
The two additional control terms in (6) are given as whereas Vb (V) is the battery voltage supplying the power
follows: circuit; N (rpm) is the rotational speed of the high-
pressure pump that is equal to N = Ne /2 with Ne being
uI (t) = KI (t) xI (t) , (12) the engine speed; Tinj (ms) is the injection time. The terms
c(N ) and d(N ), given in Appendix, are nonlinear functions
uN (t) = KN (t) sgn (ye (t)) , (13) that strongly depend on N . (See Appendix A for further
with details for the model of the residual pressure η used in the
numerical validation.)
4175
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
P, Pr [bar]
60
as a tradeoff between convergence time and reactivity of
the control action; (ii) locking of the adaptive gains was 50
50
40
5. NUMERICAL RESULTS
30
20
The NEMCSI adaptive algorithm designed in Sec. 4 has
been widely tested on the CR model (A.1) (or equivalently 10
70 80 90 100 110 120
t [s]
(A.4a)-(A.4b)) in order to prove its effectiveness in control-
ling the CR dynamics in terms of regulation, tracking and (b)
disturbance attenuation. In particular, assuming that a
Fig. 4. NEMCSI Control results for the long manoeuvre.
periodic solution of period T̄ is achieved after the time in-
Rail pressure (blue line), reference pressure (red line)
stant t⋆ when the input reference to the closed loop system
on the time interval: (a) [0; 30] (s) and (b) [62; 125] (s).
is a constant pressure Pr under the periodic disturbance
excitation η(t) (see equation (A.3) in Appendix A), we
measure quantitatively the degree of the attenuation of the closed loop (see also Fig. 3 for a graphical representation
disturbance η according to the attenuation factor defined of ∆p , ∆η and Φ).
as follows: In what follows some remarkable results are briefly de-
∆η − ∆p
Af = 100 , (19) scribed.
∆η
where ∆p = [maxΦ (p(t) − Pr ) − minΦ (p(t) − Pr )] is the • As first case we consider a long manoeuvre composed by
maximum variation of the pressure in closed loop around a sequence of pressure steps. In this case the engine speed is
the pressure working point Pr and computed over a fixed at Ne = 1000 (rpm) while the injection time for each
fixed reference pressure is listed in Tab. 1. The excellent
generic time interval Φ , [t1 ; t2 ] so that t1 > t⋆ and tracking is shown in Figs. 4a and 4b. (Notice that from 30
t2 − t1 > T̄ . Analogously, ∆η = [maxΦ (η) − minΦ (η)] is (s) to 62 (s) the reference pressure is kept constant at 30
maximum variation of the pressure disturbance due to (bar).)
the HP pump and fuel injections computed on the same
time interval. Obviously, higher Af values are obtained for For each pressure working set-point the attenuation factor
effective disturbance rejections, instead negative values of Af has been computed. Our analysis has shown that for
Af denote that the residual/ripple pressure is amplified in each constant set point Pr a satisfactory attenuation factor
4176
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
Table 1. 10
Pr = 25
Simulation Pr = 35
interval (s) 0-52 52-62.5 62.5-80 80-95 95-112.5 112.5-125 9
Pr = 45
Tinj (ms) 0 8 4 2 4 2
8 Pr = 55
Pr = 65
6 43.5 7 Pr = 75
Tinj [ms]
4 43 Pr = 85
42.5 6
2
P−15[bar], d [bar]
42
0 5
δ [%]
41.5
−2 41
4
−4 40.5
40
−6 3
39.5
1000 2000 3000 4000 5000 6000
2.65 2.7 2.75 2.8
t [s]
2.85 2.9 2.95 2.65 2.7 2.75 2.8
t [s]
2.85 2.9 2.95 Ne [rpm]
(a) (b)
Fig. 6. Injection time for each working point belonging to
8
17
P × Ne .
6
4 16
85
2
P−100[bar], d [bar]
15
0
δ [%]
80
−2 14
−4
13
−6 75
−8 12
−10
11
Af [%] 70
118.65 118.7 118.75 118.8 118.85 118.9 118.95 119 118.65 118.7 118.75 118.8 118.85 118.9 118.95 119 Pr = 25
t [s] t [s]
65 Pr = 35
(c) (d) Pr = 45
60 Pr = 55
Fig. 5. NEMCSI Control results for the long manoeuvre Pr = 65
for some steady state. Residual fuel pressure p(t) − Pr 55 Pr = 75
(blue solid line) and disturbance η acting on the plant Pr = 85
(red dashed line) when the reference pressure is: (a) 50
1000 2000 3000 4000 5000 6000
Pr = 15 (bar) and Tinj = 0 (ms); (c) Pr = 100 (bar) Ne [rpm]
and Tinj = 2 (ms); (b) control action for the case (a);
(d) control action for the case (c). Fig. 7. NEMCSI control results. Attenuation factor for
each working point belonging to P × Ne .
is attained, i.e. Af ∈ [80.92; 92.46] for the entire manoeu-
vre. To confirm the disturbance rejection performance, in out in Bursi et al. (2010), discretization implies the need
Fig. 5 detailed time histories of the residual closed loop of some bounds on the magnitude of the adaptive weights
pressure, the control action and the rejected disturbance (α and β) to avoid unwanted closed loop dynamics then,
η are shown in the case Pr = 15 (bar) (minimum refer- as consequence, the control performance deteriorates. Nev-
ence pressure over the manoeuvre) and Pr = 100 (bar) ertheless, the results here obtained with a sampling time
(maximum reference pressure over the manoeuvre). implementable in the experimental setup used in di Gaeta
• To prove the effectiveness of the NEMCSCI algorithm in et al. (2011a) are very satisfactory if we consider that not
attenuating the pressure disturbance in different working any a priori knowledge of plant dynamics have been used
conditions of reference pressure and engine speed, the for the control design. Results related to the continuous
attenuation factor (19) has been computed for each pair implementation of the control strategy are not reported
(Pr , Ne ) ∈ P × Ne with P , {25 + 10k, k = 0 . . . 6} and here for the sake of brevity.
Ne , {1000 + 500j, j = 0 . . . 10}. For each working point Details on the time history of the residual closed loop
the injection time is that to guarantee the stoichiometric pressure are given for the sake of completeness together
regime of the 2-liters GDI engine, presented in di Gaeta with the control action and the rejected disturbance η
et al. (2010, 2011b), at full load conditions (Wide Open in Fig. 8 in the case (Pr , Ne ) = (35 (bar), 2000(rpm))
Throttle, WOT). Fig. 6 and Fig. 7 show the injection (low engine speed and reference pressure) and (Pr , Ne ) =
time and the attenuation factor for each pair belonging to (65 (bar), 6000 (rpm)) (high engine speed and reference
P × Ne respectively. As it was expected, Fig. 7 shows that pressure). In the latter case the control action in Fig.
the attenuation factor is a decreasing function of the CR 8d appears less smooth than the previous cases since
pressure and engine speed. Nevertheless, the performance the sampling frequency (1 (kHz)) is comparable with the
index (19) never goes down to 55% and it remains greater highest disturbance frequency (200 (Hz)) due to injectors
than 70% independently of the pressure as far as the engine (see Appendix A) at this engine speed. Nevertheless the
speed is below 2500 (rpm). adaptive strategy still works effectively.
We remark that much better attenuation factors can be In order to stress the efficiency of the NEMCSI control
achieved if the control action (6) is not discretized, or dis- action, the controller proposed in di Gaeta et al. (2011a)
cretized with a smaller sampling time. In fact, as pointed has been tested on the same working points. That control
4177
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
12
39
8
2000
6 38 10
4
37
8 1500
2
P−35[bar], d [bar]
Tinj [ms]
36
N [rpm]
0
6
δ [%]
e
−2 35 1000
−4 4
34
−6
500
−8 33 2
−10 32
0 0
−12 0 1 2 3 4 5 0 1 2 3 4 5
31 t [s] t [s]
18.32 18.34 18.36 18.38 18.4 18.42 18.44 18.46 18.48 18.32 18.34 18.36 18.38 18.4 18.42 18.44 18.46 18.48
t [s] t [s]
(a) (b)
(a) (b)
Fig. 10. Experimental traces of (a) injection time and (b)
8
36
34
engine speed measured in a commercial ECU of a 2-
6
32
liters GDI engine.
4
2 30
P−65[bar], d [bar]
0 28
δ [%]
−2
26
−4
24 120
−6
22
−8
−10 20
100
−12 18
18.44 18.45 18.46 18.47 18.48 18.49 18.5 18.44 18.45 18.46 18.47 18.48 18.49 18.5
t [s] t [s]
80
(c) (d)
5 P = 85
r
• As last case we test here the pressure tracking perfor-
f
4178
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
50
Control, volume 3927 of Lecture Notes in Computer
40
Science, 79–92. Springer Berlin / Heidelberg.
30 Bursi, O., Stoten, D.P., Tondini, N., and Vulcan, L. (2010).
20 Stability and accuracy analysis of a discrete model
10
reference adaptive controller without and with time
delay. International Journal for Numerical Methods in
0
0 1 2 3 4 5 Engineering, 82(9), 1158–1179.
t [s]
Chatlatanagulchai, W., Aroonsrisopon, T., and Wanna-
(a) ton, K. (2009). Robust common-rail pressure control
for a diesel-dual-fuel engine using QFT-based controller.
1.5
SAE Technical Paper, (2009-01-1799).
Corno, M., Savaresi, S.M., Scattolini, R., Comignaghi, E.,
1 Sofia, M., Palma, A., and Sepe, E. (2008). Modelling,
parameter identification and dynamics analysis of a
0.5 common rail injection system for gasoline engines. In
adaptive gains
4179
Preprints of the 18th IFAC World Congress
Milano (Italy) August 28 - September 2, 2011
(a) (b)
4180