A View of Automotive Control System

A View of Automotive Control Systems Uwe Kiencke

ABSTRACT: Control of automotive vehicles and engines is a relatively new field in automatic control. Some current applications for engine control are described. Future nonlinear and time-varying automotive systems will require the development of more advanced control schemes. Representative ex- amples for plant modeling, parameter and state estimation, and adaptive control are presented.

Introduction Today’s advanced automotive systems,

such as engines, transmissions, active sus- pension systems, and brakes, are controlled by microcomputers. Introducing automatic control into these systems, one has to cope with nonlinear plant characteristics, time- varying parameters, and fast dynamics. Some variables essential for the control scheme cannot be measured at all. Thus, automotive systems are a major new challenge for control engineers.

At present, automatic control is used to enhance performance of cars currently in production. Precise control of the air-to-fuel (AIF) ratio is required to efficiently utilize catalytic converters to minimize exhaust emissions. Knock control is camed out mainly by using heuristic procedures, taking partial advantage of self-adaptive schemes. Idle speed control has been implemented with conventional proportional-integral-deriva- tives or with more advanced state-space control, additionally relying on real-time models of the intake manifold.

points lessen the prospects for finding an acceptable compromise between closed-loop stability and tracking performance. Increas- ingly integrated automotive systems will be handled by multivariable or hierarchical multilevel control approaches.

Present Control The application of electronic control of en-

gines has allowed a substantial reduction in emissions of toxic exhaust gases and improved drivability and specific fuel consumption. This section discusses control of fuel ratio, knock control, and optimal fuel consumption.

Lambda Control

Gasoline engines must get a proper mixture of air and vaporized fuel to bum efficiently. The A/F ratio is called lambda (A) and is calculated as follows, where rizL is the mass airflow into the engine, mK the metered fuel mass, and LsT the A/F mass ratio of stoichiometric mixture.

h = (ritL/rn,)/LS, (1 )

Lambda is normalized to be 1 at stoichiometric operation. While the driver regulates the amount of air, mL, going into the cylin- ders, the correct quantity of fuel, m,, must be added by the electronic control unit. Due to incomplete fuel evaporation, gas turbu-

components may be oxidized and reduced by means of a catalytic converter, which is built into the exhaust pipe system. The converter operates, however, only if the A/F ratio lambda is controlled precisely to stoichi- ometry. The average lambda offset must be kept below 0.1 percent. In the case of engine transients, short excursions may be tolerated up to 3 percent.

With the help of a so-called lambda sensor, the A/F ratio in the exhaust pipe can be measured. The sensor generates a voltage showing a steep gradient in the area of lambda equal to 1. Hamburg and Shulman [ l ) have shown that, in this application, the engine may be described by a simplified model, consisting of a delay time T,, and a lag time T,. These times vary with the engine’s operating condition in the range of delay time TL approximately equal to from 100 msec to 1 sec and lag time T, approximately equal to from 50 msec to 0.5 sec. The variation is considered in the controller by respectively adapting its parameters. The block diagram of the entire control loop is shown in Fig. 1 .

Due to nonlinearities and delays, a limit cycle of lambda near the nominal value ap- pears. If the size of control parameter T, is approximately matched to lag time T,, then the frequency of the limit cycle wx is determined exclusively by the delay time TL.

lence in the combustion chamber, and the limited time available, the burning process is nonideal, even at stoichiometric

bons, and, simultaneously, nitrogen oxide are generated in addition to nontoxic carbon dioxide and water vapor. The dangerous

wg = ?r/(2T[,) (2)

The integration time TR depends on the still Poisonous carbon monoxide, hydrocar- amplitude Of the l imit cyc1e

TR = ( 8 b / r 2 ) ( K , K l j A R ) TL (3) Further gains in performance require in- traduction of modem control theory into these applications. However, estimation procedures for state variables as well as for model parameters must be advanced to en- courage their usage in real systems. Limited performance, even of future microcontrol- lers, restricts computational complexity. Rapid plant changes to different operating

The control results of the A/F ratio lambda

Presented at the 1987 IFAC World Congress, Munich, Federal Republic of Germany, July 26- 31, 1987. and published in the Congress Proceed- ings. This work was done while the author was with Robert Bosch Corporation, Stuttgart, Federal Republic of Germany. Uwe Kiencke now works at Siemens Corporation, Siemensstr. 10, D-8400 Regensburg 1 , Federal Republic of Ger- many.

1 r--.-.----

Comparator PI-Controller I Engine Model i

L . - . - . - . i Lambda-Sensor

Fig. 1. Block diagram of lambda control [l].

0272-170818810800-001 1 $01 00 0 1988 IEEE Auguht 1988

Lambda Controller

Lambda

HC

NOX

CO

1.02c I

I I I 1 I I I I I I I I I

.:; 0.1

0.1 o?zIzzl 0 I

0 1.0 2.0 3.0 4.0 s Time

Performance of lambda control at a real engine before Fig. 2 . and after catalyst.

and of toxic exhaust emissions measured at a real engine before and after catalyst are shown in Fig. 2.

The lambda signal before the catalytic converter oscillates within the desired range. At the converter output, lambda is kept very precisely to stoichiometric 1 in stable operating points of the engine.

The performance of this control scheme has been enhanced further by the introduction of an adaptive open-loop control map [2]. If, over the vehicle’s lifetime, open-loop control electronics drift away in different operating points, closed-loop control has to compensate. The integration time TR of the controller must not be reduced below the limit given by Eq. (3), resulting in a tem- porary mismatch of A /F ratio during engine transients [Fig. 3(a)]. An adaptation of the open-loop control, derived from the closed- loop controller output signal, reduces the duration of these transient phases [Fig. 3(b)].

The correction values for every operating point of the engine are “learned” into a control map that is stored in a nonvolatile memory. Figure 3 shows the benefits of this approach for a vehicle during part of the Federal Test Procedure emission test cycle. The closed-loop controller no longer has to

compensate for the mismatches if the learning map already takes care of them.

Knock Control

In gasoline engines, an unburned part of the A/F mixture within the combustion

Lambda control, mismatch during engine transients

. . . . . . . _. .- - . - . . - Time (4

Lambda control with adaptive open-loop control map

U Vehicle speed Time + rywrV Lambda controller

(b)

Fig. 3. mance by adaptive lambda control [2].

Improvement of transient perfor-

chamber may autoinflame prematurely. This phenomenon is called knocking, because it generates resonating gas oscillations that can be heard as knocking and threaten to destroy the engine [3]. The flammability of gasoline depends on

temperature

pressure

compression ratio

time duration under stress

fuel quality

A forward-loop control cannot consider all these influences. Closed-loop knock control provides a more accurate compensation by regulating the ignition angle or the boost pressure in turbo engines [4]. Alternatively, knocking can be measured by

cylinder pressure,

engine block vibrations,

ion current through gas mixture,

light emission within the combustion chamber.

Regardless of the adopted measurement technique, knock-induced modulations of the signal are band-pass filtered. In case of knock detection, the ignition angle is retarded by AaR (Fig. 4), postponing the thermodynamic process and thereby reducing maximum cylinder pressure and temperature. Without the occurrence of knocking, the ignition angle is again gradually advanced along a slope da,ldt.

By heuristically adjusting AaR and da,/dt, the average knocking frequency and transient behavior of the control are determined. The ignition angle oscillates around an average retardation value, always approaching the knock limits under different operating conditions of the engine.

The efficiency of knock control also can be improved by adaptation. However, unlike lambda control, procedures globally effective in all the engine’s operating points have not been found as yet. In [SI, Kiencke and Cao have derived an adaptation scheme that

occurrence

Knock 1r- Time - Reiard -

Fig. 4. Knock control [4]

I‘! IEEE Control Systems Mogozine

takes into account a statistical analysis of the individual driver's habits in different operating points.

Sel f Optimizing Control of Fuel Consumption

If a gasoline engine is operated with constant output torque and speed, fuel consumption may be minimized by increasing the amount of intake air ljlL up to an optimum value. The stationary behavior of the engine thus may be characterized by a quadratic function of specific fuel consumption over air 161. The aim is to operate the engine on the lean side at minimum fuel consumption.

The optimum operating condition of the engine may be found by superimposing an orthogonal test function AmL (air modula- tion) to the intake air mL and cross-correlat- ing it with an output signal dnldt proportional to engine torque (Fig. 5). The result of the correlation contains the gradient of the quadratic function as part of the engine model.

(dnldt) dt (4)

The gradient is regulated to zero by integral control. A problem in applying this self- optimizing control has turned out to be the slow response time of engines. The intake air may be modulated no faster than about 2 Hz. If a sufficiently large correlation time

Fig. 6. Engine map for ignition angle [7].

constant T, is provided, the controller time constant goes up to = 30 sec. Therefore, applications are limited to learning control of open-loop engine maps.

Advanced Control Advanced modem control can improve

performance and provide more systematic approaches, such as

Plant modeling

Parameter and state estimation

Adaptive and robust control

Integral Control

I F( jw)l e-'"

Test Function A

AmL = AmL COS t

Sample + Hold n Delay 1

I u,(t) b r 4 1 - e e X ~ - Cross Correlation i T m P J Fig. 5. Self-optimizing control of fuel consumption [6].

Hierarchical, decentralized control

Intelligent control

Plant Modeling

It is well known that, prior to controller design, the dynamic behavior of the respec- tive process must be analyzed. This requires derivation of a model that describes process properties to sufficient accuracy. In automotive systems, the modeling job turns out to be far more difficult and time-consuming than the design of the controller itself. Some models currently in use only address static plant behavior; for example, so-called engine maps have been introduced to optimize A/F ratio and ignition angle in each operating point of the engine (Fig. 6). The ignition angle is stored for all combinations of engine speed and basic fuel-injection timing. After having determined its present operating point, the ignition controller reads out the optimum ignition angle from the map.

Boning and coworkers [7] and Tennant and coworkers [8], [9] developed a computerized procedure to get the optimum values, which is based on the method of Lagrangian mul- tipliers. By this, the engine's fuel consumption is minimized while regarding given con- traints for pollutants.

Unfortunately, such static models only partially describe transient behavior of the engine. Currently, only a few useful dynamic models exist [lo]-[15]. This is due to the problem that models tend to become very complex to match well the physical behavior of the plant; for example, an engine process is comprised of the following disciplines:

aerodynamics

thermodynamics

August 1988 13

flanic propagation

chemical reaction

mechanics

Automotive engineers have the unpleasant decision of either simplifying models for easy handling, at the risk of unduly reducing their validity. or else leaving models rather complex. being unable to parameterize them for the real world. Plants are usually extremely nonlinear. time variant, and coupled. Since control synthesis is based primarily on linear. time-invariant theory, models have to be linearized around operating points. By doing this, an additional dependence on operating-point characteristics is introduced.

One way to overcome these problems may be to combine partially dynamic models with static control maps. Kiencke and Schulz [16] have derived a model, describing the airflow intake manifold of a gasoline engine (Fig. 7 ) . For simplification, the intake manifold will be regarded as a concentrated element, where local pressure distribution and pulsa- tions are not considered. Fuel is injected on top of the intake valves of the engine, assuming the gas in the intake manifold to be ideal. The quantity of heat necessary to evaporate fuel will be provided completely by heat conduction from the engine. State transitions of the gas inside the intake manifold arc then considered as adiabatic. The energy balance describes the alteration of the specific inner energy U, inherent to the air mass rn, within the intake manifold volume V, due to the specific enthalpies h , and h, of in- and outflowing air in , and h, as

[ d ( ~ i , ~ , , ) ] / d t = m,h, - h u h , ( 5 ) Execution of differentiation and introduc-

tion of air density ps, ambient and manifold temperatures T,, and Tr, and specific heats cp and cr . results in

P ~ T , + T y & = (cp/c,3 (1/V5)

. (heTu - h o T J (6)

The ideal gas law is applied to this expres- sion, yielding

pT = (KRT,,/V,) (m, - m,T,/T,,) ( 7 )

The inflow he shall be measured, e.g. , by an airflow meter. Taylor [I71 has approxi- mated outflow ma from the manifold into the engine employing two constants K 2 and K 3 , which depend on the engine’s operating points and the exhaust gas pressure, respectively.

However, this frequently employed formulation is not accurate enough for use over the engine’s entire operating range. A reason may be the variable backflow of burned gases into the manifold through the open intake valve, which has not been considered in the preceding approach. Improved model validity can be obtained if the outflow is given by a static control map

ha = f ( k P,) (8)

which is determined by engine test-bed measurements at stationary airflow. The dynamic behavior of the resulting model

is = ( W s ) (h, - f(& PJTJT,) (9)

is dominated by the time constant

T,, = V,/(KRT,) (10)

A mechanical model, not detailed in this paper, completes the description of engine performance. The comparison of manifold pressure and engine speed shows excellent tracking between estimated and measured values (Fig. 8). The preceding model has been introduced successfully into state-space control of engine idle speed, enabling re- moval of an additional manifold pressure sensor [ 161.

R. I . Morris and coworkers [18] and T . Takahashi and coworkers [32] have taken an alternating approach by employing a black- box model. Such models may be used for

Fig. 7 . Airflow model of intake manifold [16].

-- 0 2 4 6 8 1 O c

T me

Fig. 8. Modeling of intake manifold [16]

control synthesis in some areas. However, they never reproduce physical relationships within the process, which are essential for model validation and diagnostics. Therefore, it is recommended that black-box models be applied only if a physical model cannot be obtained.

Parameter and State Estimation

Assuming the availability of a suitable model, it has to be parameterized. If processes are linear and time invariant, an ob- vious approach would be to determine parameters in an off-line configuration [19]; for example, an engine might be excited on the test bed with input test functions while mon- itoring output values. Well-known methods, such as correlation techniques, Fourier analysis, and pulse or step response, apply here. Unfortunately, off-line techniques poorly cover most nonlinear, time-variant automotive applications. Even in cases where fixed parameters might be a good start, an auto- mated offset for manufacturing tolerances and aging is required. Therefore, recursive on-line estimation procedures are favored in conjunction with adaptive control.

Out of a large variety of recursive on-line estimation methods [20], only a few are suitable for automotive applications. Contrary to chemical processes, time constants within the engine, transmission, and antiskid brake systems are orders of magnitude smaller, start- ing in the range of only milliseconds. There- fore, algorithms for parameter estimation must converge very rapidly. At the same time, closed-loop stability must be main- tained. Further complicating the situation, limited performance even of future micro- controllers constrains computational complexity of estimation algorithms.

Cao et al. have given an overview of identification methods for automotive applications [21]. A recursive least-squares (RLS) algorithm has been extended to identify

14 I€€€ Control Systems Magazine

variable parameters p. The basic scheme may be formulated as

p ( k + 1) = p(k) + r(k + 1) e(k + 1)

(11)

The gain r(k + 1) may be calculated as shown. where P ( k ) is the vector of actual input and output values.

r(k + 1) = l ( k ) P(k) /

11 + P%)l(k) P(k)l (12)

The error of real versus estimated output is

e ( k + 1) = ~ ( k + 1) - j ( k + 1) (13)

The covariance matrix is updated accord- ing to

k ( k + I ) = ~ ( k ) - r(k + 1) P r ( k ) ~ ( k )

(14)

For clarity, l ( k + 1) will be alternatively expressed as

A ( k + 1) = b(k)/[l + PT(k) l ( k ) P(k)]

(15)

which allows conversion of Eq. ( 1 1) to

p ( k + I ) = p(k) + l ( k + 1)

. P(k) e ( k + 1) (16)

The elements of the covariance matrix l ( k + 1) may be interpreted as weighting factors, determining the impact of measured values P ( k ) and error e (k + 1) on the convergence rate of estimation vector p(k). Equation (15) shows that the elements of l ( k + 1) decrease continuously during the estimation procedure. Finally, convergence rates become very low, which can be tolerated only in the case of fixed parameters.

However, the requirement was to identify time-variant parameters. Therefore, the preceding scheme must be modified-e.g., into

RLS with forgetting factor [22], [23],

RLS with moving window [24],

RLS with resetting [25].

Returning for a moment to nonrecursive least squares, the covariance matrix may be formulated as

l ( k ) = [M(k) MT(k)]-’ (17)

with M(k) being the concatenation

M(k) = tP(O), P(1), . . . , P ( k - 111

(18)

Advancing from k to k + I , the latest vec- of vector for measurement values.

tor P(k) is added to M(k), leading to

M(k + 1) MT(k + 1)

= M(k) MrW + P(k) PTW (19)

Running through the recursion steps, M(k) is continuously enlarged, resulting in de- creasing elements of l ( k ) [Fig. 9(a)]. RLS with forgetting factor introduces a factor hF into Eq. (19). Vector P within matrix M(k) is thus exponentially decreased versus time, emphasizing recent measurements [Fig. 9(b)l.

M(k + 1) Mr(k + 1)

= AFM(k) Mr(k) + P(k) P T ( k ) (20)

Equation (14) is modified by forgetting factor XF into

l ( k + 1) = (l/hF) [ l ( k ) - r(k + 1)

. P T(k) k(k)l (21)

increasing the elements of the covariance matrix l ( k + 1) over those in RLS. Usually the forgetting factor X F between 0.95 and 1 .O is chosen. However, it may be preferable to have hF depending on error e (k ) to avoid a blow-up effect [26], [27]. Fortescue et al. (221 have proposed that

XF = 1 - ( e2 (k ) /Co) /

t 1 + P r(k) l ( k ) P (41 (22)

8 Standard RLS

K = l F [

K = 2

K = 3

@RLS with moving window technique

with constant CO being determined by application events. Thus, X F is decreased below one, when parameter transients are indicated by a large error e @ ) , and is almost equal to one in steady situations with small errors.

RLS with “moving window” deletes the measurement vector P(k - rn) from matrix M(k) while introducing a new vector P ( k ) . Thus, a moving observation window on the latest rn vectors P(k) to P ( k - rn + 1) is formed [Fig. 9(c)]. In some cases, a com- bination of RLS with both forgetting factor and moving window has been considered [Fig. 9(d)].

In some automotive applications, RLS with resetting has delivered promising results. After a given number of recursion steps, the time scale is reset to its initial value. The last estimated parameter vector p ( k ) is taken over as the initial value p(0) into the next calculation interval, while the covariance matrix l ( 0 ) is reset to an initial diagonal matrix. The original RLS algorithm as stated earlier is restarted repeatedly, using Eqs. (11)-(15). A “fall-asleep effect” of plain RLS thus can be overcome efficiently without endangering stability.

Comparative results of the different approaches are shown in Fig. 10. The friction coefficient between the tire and the road surface at one of a car’s wheels is recursively

@RLS with constant forgetting factor kF = 0.5

B

@ RLS with both forgetting factor and moving window a- m rn

Fig. 9 . Recursive least-squares (RLS) estimation procedures P I ] .

August I988

whether estimated values are within their tolerated reference range. The difficulty is to find a trade-off between high sensitivity- eventually triggering false alarms-and low sensitivity-risking not detecting faults at all. As many more sophisticated electronic systems will be installed into automobiles, all disturbances. proper design of early fault-diagnosis sys-

- - ~ - ~~, T y ( t ) + y ( t ) = Ku(t) + z (23) I C 2 0 3 o tems will be a major challenge for control

The linearization of the model in operating points U, and V, leads to a first-order system with variable parameters, with output signal (speed), y( t ) the state variable, and control signal (throttle angle) U ( [ ) , proportional factor K , time constant T, and z summarizing

' t -

y ( t ) = vw - v, (24) Time engineers in the future.

Fig. IO . Recursive least-squares estimation of the friction coefficient between tire and road surface [21].

estimated during an antiskid braking procedure. After about half a second, the condition of the road surface changes almost step- wise. RLS with resetting is tracking this with a lag time of under 100 msec.

State-variable estimation also is required frequently in automotive control systems. In an example presented earlier [ 161, manifold pressure was obtained from a model rather than an additional sensor. In state-space control, only part of the state variables may be accessible. State-variable estimation allows for substantial cost savings by replacing measurement devices. The usual way to handle this task is to apply a Kalman filter [28]- 1301 or Luenberger observer [31] . Ta- kahashi and coworkers [32] have employed an observer for idle speed control. There are a number of problems to overcome in real application:

In order to estimate state variables, the structure and parameters of the process model have to be accurately known. Ob- viously, this precondition is not always met.

In Kalman filtering, statistical characteristics of input and measurement noise must be known. However, in some applications, statistical data are assumed arbi- trarily, without any reference to the real situation.

As a consequence, state-variable estimates differ from real values, making such approaches less attractive. Proposals to overcome thesc problems are adaptive Kalman filtering 1331, [34] or heuristic handling of statistical data [35] , [36] . However, results are not always satisfactory. Further progress is still necessary in this area.

Anothcr application of estimation procedures in automobiles is early diagnosis of emerging faults [ 3 7 ] , [38]. Based on a val- idated model, system parameters and/or state variables are estimated continuously. An emerging fault can be detected by checking

Adaptive and Robust Control

Assuming the existence of a model with sufficiently accurate parameters, adaptive control schemes can be applied. Robust control is another way to regulate plants with only slight parameter variations, if a controller with constant parameters is acceptable [ W .

Cao et al. [21] have investigated an adaptive cruise control for vehicles, which re- lieves the driver from regulation at a given speed. For different engine types, transmission gears, and loading conditions, plant parameters vary significantly. Characteristics of engine output torque and air drag are nonlinear. Conventional proportional-integral control performs rather poorly in this situation.

A model for cruise control is shown in Fig. 1 1. Here, KE is the amplification factor of the engine torque map determined by test- bed measurements, i, and iA the transmission ratios of the gearbox and the drivetrain, 7, the transmission efficiency, and r the wheel radius. The subtraction of different load

u(t) = U ( t ) - U, (25)

K = KE(i,iATdr)/

( P L c,, AV,) (26)

T = T,J(PL~, .AVO) (27)

After introduction of variables a , and b, , this can be transferred into a time-discrete description.

~ ( k + 1 ) = a , y ( k ) + h , u ( k ) (28)

a , = exp ( - T,a,np~e/T) (2%

b , V ( l - U , )

+ ~ ( 1 - ai)/[u(k)l (30)

The disturbances are contained in parameter 6,. For control, a one-step-ahead controller that minimizes the cost function J is adopted because of its extraordinary simplicity.

J = p [ w ( k + 1) - y ( k + 1)12 + qu2(k)

(31) forces from the driving force Fd yields the accelerating force Fb, which is integrated into

portional to the effective mass of the vehicle, which also covers rotational movements of the drivetrain. The air drag force is calculated from the vehicle speed v, the wind velocity v,, the air density pL, the drag coefficient c,, and the sectional area A .

The control law can be stated as

'peed. vehicle speed v , The time constant T, is pro- where w(k + is the reference vehicle

~ ( 4 = [pb , / (q + pb?)l

~k + 1) - a ,y(k) i (32)

Model parameters a , and b, are estimated

Engine Torque Transmission Eff. Vehicle Mass Man + Drivetrain

Vehicle r

I I - - I - I I

Control Signal U

Gradieht Rolling Wind' Resistance Friction velocity vW - "

Disturbances

Fig. 11. Plant model for cruise control [21].

16 I E E E Control Systems Magazine

Loading Unloading I -P--L-- Ampl~ficalian Factor

Vehicle Speed

v ~ - - -. -I-L---_

6 12 18 24 30

Time

Fig. 12. cation to different loading situations in cruise control 1211.

Adaptation of controller amplifi-

continuously, e .g . , by RLS, and then en- tered into Eq. (32).

Figure 12 shows a representative result from field experiments in the car. The controller’s amplification factor

(33)

readily adapts to transients between loading and unloading, simulated by pulling manual brakes. Vehicle speed deviates only slightly in these situations, proving good practical performance of adaptive control. Other ad- vantages are as follows:

covering a broad range of vehicle types

significant reduction of application work,

self-tuning against aging during the car’s

with only one basic controller,

and

lifetime.

Hierrirchiccrl, Decentralized Control

In advanced cars, a number of electronic control subsystems are implemented, such as engine control, transmission control, performance control, antiskid braking control, sus- pension control, etc. Until recently, these systems operated mostly on a stand-alone basis. Since all subsystems are coupled via the car itself, the control activities in one individual subsystem can generate undesired cross influences within the others. To overcome such effects, an additional optimization control layer must be introduced, operating in addition to already existing subsystems. This leads to a distributed mul- timicrocomputer system in the car.

For implementation, Robert Bosch GmbH has announced a new automotive communication link called “controller area network” (CAN), which provides not only communication between subsystems but also supports parallel processing in distributed controllers 1401, [41].

CAN provides basic mechanisms for process synchronization and consistent data

handling. User programs in the individual controllers are efficiently unloaded from the communication job itself. For each message, a communication object is created, which consists of the following:

Identifier, naming and routing the mes-

Control segment, containing all transfer

Data segment, ranging from 0 to 8 bytes.

User programs may access communication objects in the communication buffer as in any ordinary random access memory at pre- defined locations. Thus, a virtual communication channel is provided for each individual message on the application layer, whereas network control is hidden from the user. Built-in CAN features include:

error detection,

error correction,

failure localization, and

automatic truncation of faulty nodes

Time for recovery from error is very short, with 25 bit times only.

The following example for an integrating control system makes use of the preceding communication type (Fig. 13). There are two

sage;

control information; and

t

separate electronic subsystems for engine and transmission control. Under certain driving conditions, the entire drivetrain might jerk. For relief, an additional control hierarchy is introduced, which employs engine load and engine, drivetrain, and vehicle speeds as in- puts, and superimposes a correction on engine performance. Control programs are lo- calized both in the engine and transmission subsystems. Measurement values and control signals are transmitted via the serial communication link.

The advantage of decentralized control is an improved availability of the overall system in case of failures; for example, in case of a failure in the transmission control computer, the transmission shifts into the highest gear. The car is still operational although with degraded functionality. Contrary to a centralized arrangement, engine control is not affected at all.

Intelligent Control

Controller synthesis as discussed thus far in this paper has been based on analytical models. Design methods have been aimed at

optimization of a cost function,

guarantee of stability,

preassignment of dynamic behavior, etc.

”!

:!

:!

:!

i ! i ! i !

Correction AU I I +... Engine Load, Speed

+.- Transmission Gear, Performance Signal

Fig. 13. Distributed vehicle control with parallel controllers 1401, [411.

17

As mentioned earlier, models of automotive systems sometimes may be difficult to derive or complicated to handle. In such cases, intelligent control might be an alter- nate method of increasing control efficiency. Learning AIF ratio control, discussed earlier, is considered to be a first pragmatic step in this direction.

In the future, such approaches must be developed more systematically, making use of pattern recognition [42], [43], expert systems [44], and fuzzy control [45], [46]. It is hoped that these methods will bridge the gap between traditional control and data-base orientation in the field of artificial intelli- gence.

Conclusion In order to improve performance of auto-

matic control in automotive systems, more advanced control schemes involving modeling, parameter and state estimation, and adaptation will be applied. An overall optimization layer combining all electronic systems will further improve driving safety and economy. To handle nonlinear and time- varying automotive systems, additional progress in control engineering is needed, especially in the areas of state-variable estimation and intelligent control.

Acknowledgment The author wishes to thank Dr. C. T . Cao

for his assistance and valuable contributions to this paper.

111

I21

I31

141

PI

161

[71

18

References R. Hamburg and M. A. Shulman, “Closed- Loop AIF Control Model for Internal Com- bustion Engines,” SAE Paper 800826, 1980. G. Felger. 0. Glockler, H. Kauff, U. Kiencke, H. Knapp, and H. Stocker, “Reg- ulating Device for a Fuel Metering Sys- tem,” U.S. Patent 4,440,131, 1982. G. Woschni, “Verbrennungsmotoren,” Lecture Manuscript, Institut fur Verbren- nungskrdftmaschinen und Kraftfahrzeuge, Universitat Munchen, Federal Republic of Germany, 1980. Automotive Handbook, Robert Bosch GmbH, Stuttgart, Federal Republic of Ger- many. 2nd Ed., pp. 396-409, 1986. U. Kiencke and C. T. Cao, “Regelver- fdhren in der Elektronischen Motorsteu- erung.“ Part 2, Automobil-Industrie 1/88, 1988. P. H. Schweitzer, C. Volz, and F. Deluca, “Control System to Optimize Engine Power,” SAE Paper 660022, 1966. B. Bdning, W. Richter, W. Tuleweit, and B. Zeilinger, “Systematische Optimierung

von Kennfeldern fur Gemischbildung und Zundung bei Ottomotoren.” Motortech- nische Zeitschrif, vol. 44, pp. 53-58, 1983. 1. A. Tennant, “Computer-Aided Proce- dures for Optimization of Engine Con- trols,” Use of Computers in the Automotive Industry, International Symposium on Au- tomotive Technology and Automation, To- rino, Italy, vol. 2, 1980. 1. A. Tennant, A. I. Cohen, H. S. Rao, and J . D. Powell, “Computer-Aided Procedure for Optimization of Engine Control,” Inr. J . Vehicle Design, vol. SP4, pp. 1-12, 1983. C. F. Aquino, “Transient A/F Control Characteristics of the 5 Liter Central Fuel Injection Engine,” SAE Paper 810494, 1981. C. F. Aquino, “Steady-State and Transient A/F-Control Requirements for Cold Oper- ation of a 1.6 Liter Engine with Single-Point Fuel Injection,” SAE Paper 850509, 1985. D. 1. Dobner, “A Mathematical Engine Model for Development of Dynamic Engine Control,” SAE Paper 800054, 1980. D. 1. Dobner, “Dynamic Engine Model for Control Development-Part I: Non-Linear and Linear Model Formulation,” Int. J . Ve- hicle Design, vol. SP4, pp. 54-74, 1983. B. K. Powell and C. F. Aquino, “Stoichio- metric Air-Fuel Ratio Control Analysis,” SAE Paper 810274, 1981. 1. D. Powell, “A Review of IC Engine Models for Control Systems Design,” Proc. IFAC World Cong. Auto. Contr., vol. 3 , July 1987. U. Kiencke and A. Schulz, “Entwurfeines Zustandsreglers fur die Leerlaufregelung eines Otto-Motors,” VDI-Berichte 612, pp. 93-108, 1986. C. F. Taylor, The Internal Combustion En- gine in Theoqi and Practice, vol. 1, Cam- bridge, MA: MIT Press, 1966. R. I . Moms, R. H. Borchert, M. V. War- lick, and H. G. Hopkins, “Spark Ignition Engine Model Building-An Identification Approach to Throttle-Torque Response,” Int. J . Vehicle Design, vol. SP4, pp. 89- 101, 1983. W. Reichelt, “Identifikationsmethoden fur die Fahrdynamik,” Automobiltechnische Zeitschrif, vol. 86, no. 9, pp. 391-397, 1984. L. Ljung and T. Soderstrom, Theory and Practice of Recursive Identijication, Cam- bridge, MA: MIT Press, 1983. C. T. Cao, H. Janetzke, and A. Schulz, “Einsatz von Identifikationsverfahren im Kraftfahrzeugbereich, Automatisierungs- technik, vol. 34, no. 9, pp. 339-346, 1986. T. R. Fortescue, L. S . Kershenbaum, and B. E. Ydstie, “Implementation of Self- Tuning Regulators with Variable Forgetting Factors,” Automutica, vol. 17, no. 6, pp. 831-835, 1981. P. E. Wellstead and S . P. Sanoff, “Ex- tended Self-Tuning Algorithm,” Int. J . Vehicle Design, vol. SP4, pp. 1-12, 1981.

P. A. Janakiraman and S. Renaganathan, “Application of a New Identification Pro- cedure for Determining the Weight in a Pas- sive Oscillatory System,” Regelungstech- nik, vol. 31, no. 1, pp. 19-25, 1983. X. Y . Xie and R. 1. Evans, “Discrete-Time Adaptive Control for Deterministic Time- Varying Systems,’’ Automatica, vol. 20, no. 3, pp. 309-319, 1984. R. Isermann, “Parameter Adaptive Control Algorithms-A Tutorial,” Automutica, vol. 18, no. 5, pp. 513-528, 1982. R. E. Kalmann, “A New Approach to Lin- ear Filtering and Prediction Problems,” ASME J . Basic Eng. , vol. 82, pp. 34-45, 1960. B. Wittenmark and K. J . Astrom, “Practice Issues in the Implementation of Self-Tuning Control,” Auromarica, vol. 20, no. 5 , pp. 595-605, 1984. W. F. Powers, B. K. Powell, and G. P. Lawson, “Application of Optimal Control and Kalman Filtering to Automotive Sys- tems,” Int. J . Vehicle Design, vol. SP4. pp. 39-53, 1983. K. C. Shet and B. V. Rao, “An Overview of the Kalman Algorithm,” Int. J . Elec- tron., vol. 59, no. 6, pp. 657-665. 1985. D. G . Luenberger, “Observers for Multi- variable Systems,” IEEE Trans. Auto. Contr., vol. AC-11, pp. 190-197. 1966. T. Takahashi, T. Ueno, A. Yamamoto, and H. Sanbuichi, “A Simple Engine Model for Idle Speed Control,” SAE Paper 850291, 1985. Y. M. El-Fattah, “Recursive Self-Tuning Algorithm for Adaptive Kalman Filtering,” IEE Proc., vol. 130, no. 6, pp. 341-344, 1983. A. H. Jazwinski, Stochastic. Processes and Filtering Theory, New York: Academic Press, Chap. 8, 1970. S. C. Iglehart and C. T. Leondes. “Esti- mation of a Dispersion Parameter in Dis- crete Kalman Filtering,” IEEE Trans. Auto. Contr., vol. 19, pp. 262-263, 1974. A. L. C. Quigley, “An Approach to the Control of Divergence in Kalman Filter Al- gorithms,” Int. J . Contr., vol. 17, pp. 741- 746, 1973. A. A. Fahs, ‘‘Fault Analysis in Dynamic Systems with Application to Reprocating Combustion Engines,” Ph.D. thesis, Wayne State University, Detroit, MI, 1985. R. Isermann, “Diagnoseinethoden mit Mo- dellbildung,” Verfahren und Systeme zur Technischen Fehlerdiagnose, Conference Report Langen, D. BarschdorlT, Ed., pp. 167-190, 1986. J . Ackermann, “Parameter Space Design of Robust Control Systems,” IEEE Trans. Auto. Contr., vol. AC-25. no. 6, 1980. U. Kiencke, S. Dais, and M. Litschel, “Automotive Serial Controller Area Net- work,” SAE Paper 860391. 1986. U. Kiencke, “Serial Communication Net- work Among Distributed Microcontrol- lers,” Wescon 1986, Paper 3012, 1986.

I E E E Control Systenis Mogorine

[42] E. H . Bristol, “Pattern Recognition: An Al- ternative to Parameter Identification in Adaptive Control,” Auromarica. vol. 13, pp. 197-202, 1977.

1431 T. W. Krans and T. J . Myron, “Self-Tun- ing PID Controller Uses Pattern Recogni- tion Approach.“ Contr. Eng., pp. 106-1 11, June 1984. J . H. Taylorand D. K. Frederick, “An Ex- pen System Architecture for Computer- Aided Control Engineering,” fEEE Proc., vol. 72, no. 12, pp. 1795-1805, 1984. M . Maeda and S . Murakami, “Speed Con- trol of Automobile by Fuzzy Logic Con- troller,” Trans. Soc. Insrru. Conrr. Eng. (Jupan), vol. 21, no. 9, pp. 984-989, 1985.

1461 M . Sugeno, “An Introductory Survey of Fuzzy Control,” fnfor. Sci., vol. 36, pp. 59-83. 1085.

1441

[45]

Uwe Kiencke was born in in Stuttgart, where he was involved in the dd- 1943 in Oldenburg, Fed- vanced development of dutomotive control sys- era1 Republic of Ger- tems and digital LSI circuits such as an automotive many He received a de- real-time microcontroller He headed the team that gree in communications developed the senal automotive cornmumcation engineenng from Karls- network “controller area network ” In 1988, he mhe University in 1967 joined Siemens Corporation in Regensburg and IS

and a doctoral degree in currently Department Manager for design of the control engineenng from electronic engine control unit He lectures at Kai- Braunschweig University serslautern University on “Control and Signal in 1972 From 1972 to Processing in Automotive Systems ” 1987, he was employed at Robert Bosch Corporation

1988 CDC Arrangements are proceeding for the 1988

IEEE Conference on Decision and Control (CDC) to be held Wednesday through Fri- day, December 7-9, 1988, at the Hyatt Re- gency Austin on Town Lake, Austin, Texas. As usual. the conference will be conducted in cooperation with the Society for Industrial and Applied Mathematics and the Opera- tions Research Society of America.

In the accompanying photograph, General Chairman Mike Polis proposes a toast to the success of the CDC, which will feature in- teresting technical presentations in an inter- esting setting. For those who appreciate elaborate architecture, also pictured is a his- toric home in the Bremond block of down- town Austin.

The conference emphasizes all aspects of the theory and applications of systems involving decision, control, optimization, and adaptation. For further information, contact the General Chairman:

M. P. Polis Dept. of Electrical and Computer

Wayne State University Detroit, MI 48202 Phone: (3 13) 577-3920

Engineering

19

A View of Automotive Control System

Documents

Transcript of A View of Automotive Control System