Sunteți pe pagina 1din 6

Proceedings of the 2006 American Control Conference Minneapolis, Minnesota, USA, June 14-16, 2006

ThB14.4

A Model-based SI Engine Air Fuel Ratio Controller


Kenneth R. Muske Department of Chemical Engineering, Villanova University, Villanova, PA 19085 James C. Peyton Jones Department of Electrical and Computer Engineering, Villanova University, Villanova, PA 19085
Abstract A novel linear state-space model predictive controller for SI engine air fuel ratio control is presented and demonstrated over a range of engine operation. The linear model-based controller is an analytical controller that does not require online optimization. Time-varying delay compensation is adapted based on the measured engine speed. A Kalman lter is used to estimate the model and unmeasured disturbance states. 1. Introduction An essential component in the reduction of automotive tail-pipe emissions is the use of three-way automotive catalysts in exhaust after-treatment systems. A three-way automotive catalyst reduces engine emissions by both oxidizing unburned hydrocarbons and carbon monoxide and reducing nitrogen oxides contained in the precatalyst engine exhaust. In order for the threeway catalyst system to catalyze both oxidation and reduction reactions, the engine must be cycled between lean (excess oxygen) and rich (excess fuel) operation. When the engine is operated lean, the excess oxygen in the pre-catalyst exhaust gas is stored in the catalyst through chemisorption with the cerium oxides contained in the catalyst preventing lean (nitrogen oxide) tailpipe emissions. When the engine is operated rich, oxygen is released from the catalyst preventing rich tailpipe emissions by oxidizing the hydrocarbons and carbon monoxide. A common operating practice for three-way catalyst systems is to cycle the pre-catalyst air fuel ratio across stoichiometric at a frequency determined during engine calibration or obtained through relay feedback from a relay type postcatalyst heated exhaust gas oxygen (HEGO) sen1-4244-0210-7/06/$20.00 2006 IEEE 3284

sor [1]. A more sophisticated control strategy is to maintain the catalyst oxygen level near the stoichiometric equilibrium state in order to maximize the time required to reach either rich or lean tailpipe emission breakthrough when subject to disturbances. Linear control strategies of this type have been proposed based on H [2], LQR [3], and IMC [4] design methods. Nonlinear model-based control formulations following this strategy include [5], [6], and [7]. A key requirement for these control strategies is good servo and regulatory control of the engine air fuel ratio. This loop is typically the secondary controller for the primary catalyst emission controller within a cascade control structure. Poor control by this secondary loop will result in poor primary loop control resulting in increased tailpipe emissions. The engine air fuel ratio is controlled by manipulating the fuel injector pulse width. This is the length of time the fuel injectors are opened to supply fuel to the engine. In this work, the air fuel ratio is measured by a wide-ranging universal exhaust gas oxygen (UEGO) sensor upstream of the catalytic converter. The diculties in achieving tight air fuel ratio control are the rather long (in comparison to the dominant time constant) and time-varying nature (due to changing engine speed) of the time delay in the pre-catalyst UEGO sensor response. Previous work in linear air fuel ratio controllers has considered LQG [8], H [9], Smith predictor [10], and adaptive [11] approaches. Nonlinear controllers include sliding mode control [12], [13], and neural network approaches [14], [15], [16]. In this work, we propose an adaptive, linear, state-space, analytical model predictive control formulation for air fuel ratio control. Analytical model predictive controllers have a closed-form

solution that does not require online numerical optimization [17]. This aspect is critical for an air fuel ratio controller. Time delay compensation is adapted based on the measured engine speed. The result is a model predictive control algorithm with time-varying time delay compensation that can easily execute within the time scale necessary for air fuel ratio control using a standard engine control module (ECM). 2. Process Model Air fuel ratio is dened as the ratio of the air mass ow rate to the fuel mass ow rate O2 1 m (1) = K m f where = 1 represents a stoichiometric air fuel ratio, m O2 is the inlet oxygen mass ow rate to the engine, m f is the fuel mass ow rate, and K is the stoichiometric air fuel ratio times the oxygen mass fraction of air. Because we are manipulating the fuel ow rate, it is more convenient to express the exhaust gas composition in terms of the equivalence ratio dened as = 1/ (2) When > 1, the mixture is rich and the corresponding pre-catalyst engine exhaust is decient of oxygen. When < 1, the mixture is lean and the corresponding pre-catalyst engine exhaust gas contains excess oxygen. Although there is signicant complexity in the system dynamics due to the eects of fuel puddling, manifold wall wetting, and the intake manifold hydrodynamics [18], the predominant dynamic characteristic of the UEGO sensor response to a change in the fuel injector pulse
1.03 1.02 1.01 1 0.99 Normalized Base Fuel

width closely resembles a FOPDT response. Figure 1 shows the response to step changes in the base fuel rate (injector pulse width) for a 2L Ford I-4 engine at 1800 rpm. The response for 3500 rpm is shown in Figure 2. As shown in the gures, the response is closely approximated by a FOPDT process with time-varying dynamics, a time delay on the order of the dominant time constant, and a noisy sensor measurement. The amount of fuel injected into the intake manifold is roughly proportional to the pulse width of the fuel injectors. The engine control module (ECM) computes a base fuel rate intended to maintain stoichiometric combustion based on an estimate of the engine air ow. Feedforward adjustments to the base fuel rate are also made by the ECM to account for measured disturbances. The model-based controller in this work maintains the pre-catalyst exhaust gas equivalence ratio at a specied target trajectory around stoichiometry by manipulating a multiplier to the base fuel ow rate as computed by the ECM. Unlike many of the previously cited references, we do not attempt to determine the base fuel ow rate. The intent of this controller is to perform feedback control by correcting the base fuel ow rate determined by the production engine management system and not to replicate these calculations. To illustrate this approach, we choose to model the system as a rst order linear process xk+1 = Axk + B ukd y = Cx (3) (4)

where the model states include a rst order contribution from the base fuel multiplier u, with a
1.05 1.04 1.03 1.02 1.01 Normalized Base Fuel

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Time (sec) 5 5.5 6 6.5 7 7.5 8 8.5

0.98 0.97 0.96 0.95 0.94 0.93

1 0.99 0.98 0.97 0.96 0.95 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 Time (sec) 5 5.5 6 6.5 7 7.5 8 8.5

Figure 1: Sensor step response at 1800 rpm.

Figure 2: Sensor step response at 3500 rpm.

3285

time delay of d sample periods, and a constant output step disturbance state p x= x p , A= a 0 0 1 , B= b 0

and the output is determined with C = [ 1 1 ]. The delay time in sample periods is computed from the number of engine revolutions required for the exhaust gas to travel from the cylinders to the UEGO sensor and the current engine speed d = 60 N (R) R t (5)

The model predictive control algorithm determines a constant base fuel multiplier at each sample period that results in the modelpredicted equivalence ratio attaining a reference target at a specied future time. The specied future time is a tuning parameter referred to as the coincidence time. The reference target is obtained by dening a rst-order reference trajectory between the current equivalence ratio setpoint and the model predicted equivalence ratio after the time delay. The model predicted state after the time delay is k+d1|k = Ad1 x k |k + x
d1 j =1

where R is the current engine speed in rpm, t is the sample period in sec, and N (R) is the number of engine revolutions which may be a function of engine speed. Implementation of the timevarying delay is through an array of past input moves and the corresponding number of engine revolutions since that input was applied. The number of engine revolutions is updated at each sample period under the assumption that the engine speed was constant at the current measured value over the previous sample period. When the number of engine revolutions matches the number required to reach the sensor, the corresponding input move is used in the model prediction. 3. Model Predictive Controller The model predictive controller in this work is a linear state-space implementation of [19] that incorporates a time-varying deadtime. A Kalman lter is used to estimate the current model and unmeasured disturbance state of the system. The covariance matrices used to compute the lter gain are determined from engine data using the technique discussed in [20]. This technique allows a more computationally ecient output step disturbance model to be employed without the performance penalty associated with output step disturbance models if the true disturbance enters through the state of the system. The model state at the current sample period is k 1 |k 1 + B u k d k |k 1 = A x x (6) k|k1 + L(zk C x k|k1 ) (7) k |k = x x where d is the current delay time, zk is the current equivalence ratio measurement, and L is the lter gain determined on a delay-free basis.
3286

Aj 1 B ukj

(8)

where ukj are the stored past inputs at the previous sample periods. The model predicted state at the coincidence time c sample periods in the future is computed assuming that the input at the current sample period uk remains constant. k+d1|k + k+c|k = Acd+1 x x k + c|k y k + c|k = C x
c d

Aj 1 B uk

j =0

(9) The reference trajectory is a rst-order dynamic response between the predicted equivalence ratio after the time delay and the current setpoint. The reference target c sample periods ref in the future, yk +c , is determined as follows k+d1|k , yk+d1|k = C x
ref yk +c

= e

3c t Tc

(10) (11)

= (1

sp )yk

yk+d1|k

sp is the desired equivalence ratio setwhere yk point, ct is the coincidence time c sample periods in the future, and Tc is the closed-loop response time. The closed-loop response time is the tuning parameter that determines the speed of response of the controller. The single future control move is determined by setting yk+c|k = ref yk +c and solving for the constant base fuel rate multiplier uk .

uk =

sp k+d1|k (1 ) yk + C CAcd+1 x

c d

Aj B (12)

j =0

1.06 1.05 1.04 1.03 1.02 1.01

Target

1.06 1.05 1.04 1.03 1.02 1.01

Target

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

1 0.99 0.98 0.97 0.96 0.95 0.94

1 0.99 0.98 0.97 0.96 0.95 0.94 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

1.11 1.1 1.09 1.08

1.13 1.12 1.11 1.1

Injector Pulse Width Multiplier

1.07 1.06 1.05 1.04 1.03 1.02 1.01 1 0.99 0.98 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

Injector Pulse Width Multiplier

1.09 1.08 1.07 1.06 1.05 1.04 1.03 1.02 1.01 1 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

Figure 3: 1000 RPM and 30% throttle.

Figure 5: 1000 RPM and 70% throttle.

4. Controller Example We demonstrate the controller performance using a nonlinear mean value engine model similar to that presented in [18]. A series of rst order dynamic models, in which K is the gain and

is the time constant, were identied using this mean value model at various engine operating conditions as shown in Table 1. The time delay ranged from 0.3 sec at 1000 rpm to 0.08 sec at 4000 rpm. Based on this analysis, the rst order model with K = 1.1 and = 0.28 is employed.
1.05 1.04 1.03 1.02 1.01 Target

1.05 1.04 1.03 1.02 1.01

Target

1 0.99 0.98 0.97 0.96 0.95 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

1 0.99 0.98 0.97 0.96 0.95 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

1.05 1.04 1.03

1.18 1.17 1.16

Injector Pulse Width Multiplier

1.02 1.01 1 0.99 0.98 0.97

Injector Pulse Width Multiplier


0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

1.15 1.14 1.13 1.12 1.11 1.1 1.09

0.96 0.95

1.08 1.07 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

Figure 4: 2500 RPM and 30% throttle.

Figure 6: 2500 RPM and 70% throttle.

3287

1.06 1.05 1.04

Target

1.05 1.04 1.03

Target

1.03 1.02 1.02 1.01 1.01

1 0.99 0.98

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

1 0.99 0.98

0.97 0.96 0.95 0.94 0.97 0.96 0.95 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

0.92 0.91

1.23 1.22 1.21

0.9 1.2

Injector Pulse Width Multiplier

Injector Pulse Width Multiplier


0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

0.89 0.88 0.87 0.86 0.85 0.84 0.83

1.19 1.18 1.17 1.16 1.15 1.14 1.13

0.82 1.12 0.81 1.11 0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5 5.5 Time (sec) 6 6.5 7 7.5 8 8.5 9 9.5 10

Figure 7: 4000 RPM and 30% throttle.

Figure 8: 4000 RPM and 70% throttle.

Figures 3 through 8 present the closed-loop controlled and manipulated variable servo response for a series of step changes to the setpoint. The controller sample period is t = 0.1 sec and the controller tuning parameters are c = 6 and Tc = 1.8 sec. As shown in these gures, the controller provides good closed-loop control over the range of engine operating conditions. Although a rst order model is used in the controller, the mean value model dynamics are second order due to the combined contribution of rst order wall wetting/fuel puddling and rst order air ow dynamics. The engine data in Figures 1 and 2 suggests that one of these dynamic lags is dominant. Throttle Position 5% 30% 70% 1000 rpm K = 1.17 = 0.31 K = 1.08 = 0.31 K = 1.08 = 0.31 2500 rpm K = 1.20 = 0.28 K = 1.10 = 0.28 K = 1.09 = 0.29 4000 rpm K = 1.12 = 0.26 K = 1.10 = 0.27

In addition to the mismatch between the actual and controller model order, fractional delay times are an additional source of model error in this application. Because the discrete control algorithm implicitly assumes that the time delay is an integer multiple of the sample period, the fractional part of the delay time must either be neglected or treated as a full sample period. This eect can be reduced with a smaller sample period, however, there is a practical minimum limit to a sample period that can be implemented. 5. Conclusions We have presented an analytical model predictive controller for SI engine air fuel ratio control. The time-varying delays inherent in this system are accounted for by adapting the time delay in the model based on the engine speed. The results of this study suggest that good performance is possible over a wide range of engine operation using a single linear model. Future work includes on-line implementation of this algorithm along with the possibility of adapting the model dynamics in addition to the time delay. Acknowledgments Support for this work from the National Science Foundation under grant CTS-0215920, Ford Mo3288

Table 1: First order dynamic response models.

tor Company, Johnson Matthey, and ExxonMobil is gratefully acknowledged. References [1] P. Eastwood. Exhaust Gas Aftertreatment. Research Studies Press, Baldock, England, 2000. [2] E. Shafai, C. Roduner, and H. Geering. Indirect adaptive control of a three-way catalyst. SAE Paper 961038, SAE World Congress, 1996. [3] A. Ohata, M. Ohasi, M. Nasu, and T. Inoue. Model based air fuel ratio control for reducing exhaust gas emmissions. SAE Paper 950075, SAE World Congress, 1995. [4] M. Balenovic, A. Backx, and J. Hoebink. On a model-based control of a three-way catalytic converter. SAE Paper 2001-01-0937, SAE World Congress, 2001. [5] G. Fiengo, J. A. Cook, and J. W. Grizzle. Fore-Aft oxygen storage control. In Proceedings of the 2002 American Control Conference, pages 14011406, 2002. [6] M. Balenovic, J. Edwards, and T. Backx. Vehicle application of model-based catalyst control. In Proceedings of the 1st IFAC Symposium on on Advances in Automotive Control, pages 351356, 2004. [7] K. Muske and J. Peyton Jones. Multiobjective model-based control for an automotive catalyst. J. Proc. Control, 16:2735, 2006. [8] J. Powell, N. Fekete, and C-F. Chang. Observer-based air-fuel ratio control. IEEE Control Systems, 18:7281, 1998. [9] L. Mianzo, H. Peng, and I. Haskara. Transient air-fuel ratio H preview control of a drive-by-wire internal combustion engine. In Proceedings of the 2001 American Control Conference, pages 28672871, 2001. [10] S. Nakagawa, K. Katogi, and M. Oosuga. A new air-fuel ratio feed back control for ULEV/SULEV statndard. SAE Paper 2002-01-0194, SAE World Congress, 2002.
3289

[11] R. Turin and H. Geering. Model-reference adaptive A/F-ratio control in an SI engine based on Kalman-ltering techniques. In Proceedings of the 1995 American Control Conference, pages 40824090, 1995. [12] J. Pieper and R. Mehrotra. Air/fuel ratio control using sliding mode methods. In Proceedings of the 1999 American Control Conference, pages 10271031, 1999. [13] J. Souder and J. Hedrick. Adaptive sliding mode control of air-fuel ratio in internal combustion engines. Int. J. Robust Nonlin. Cont., 14:525541, 2004. [14] B. Lennox, G. Montague, A. Frith, and A. Beaumont. Non-linear model-based predictive control of gasoline engine air-fuel ratio. Trans. Int. Meas. Cont., 20:103112, 1998. [15] M. Won, S. Choi, and J. Hedrick. Air-to-fuel ratio control of spark ignition engines using Guassian network sliding control. IEEE Trans. Cont. Sys. Tech., 6:678687, 1998. [16] C. Beltrami, Y. Chamaillard, G. Miller, P. Higelin, and G. Bloch. AFR control in SI engine with neural prediction of clyinder air mass. In Proceedings of the 2003 American Control Conference, pages 14041409, 2003. [17] M. Masoud and K. R. Muske. Analytical model predictive control. In Nonlinear Model Predictive Control, volume 26 of Progress in Systems and Control Theory, pages 163179. BirkhauserVerlag, 2000. [18] C. Ferguson. Internal Combustion Engines. Wiley, New York, 1986. [19] J. Richalet, A. Rault, J. Testud, and J. Papon. Model predictive heuristic control: Applications to industrial processes. Automatica, 14:413428, 1978. [20] B. Odelson, M. Rajamani, and J. Rawlings. A new autocovariance least-squares method for estimating noise covariances. Automatica, 42:303308, 2006.

S-ar putea să vă placă și