- Email: [email protected]
The exploitation of neural networks in automotive engine management systems
Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
www.elsevier.com/locate/engappai
The exploitation of neural networks in automotive engine management systems P.J. Shayler a,*, M. Goodman a, T. Ma b a
University of Nottingham, Nottingham, NG7 2RD, UK Ford Motor Company Research and Engineering Centre, Essex, SS15 6EE, UK
b
Received 1 October 1998; accepted 1 October 1999
Abstract The use of electronic engine control systems on spark ignition engines has enabled a high degree of performance optimisation to be achieved. The range of functions performed by these systems, and the level of performance demanded, is rising and thus so are development times and costs. Neural networks have attracted attention as having the potential to simplify software development and improve the performance of this software. The scope and nature of possible applications is described. In particular, the pattern recognition and classi®cation abilities of networks are applied to crankshaft speed ¯uctuation data for engine-fault diagnosis, and multidimensional mapping capabilities are investigated as an alternative to large `lookup' tables and calibration functions. # 2000 Elsevier Science Ltd. All rights reserved. Keywords: Automotive engines; Calibration; Engine control; Fault diagnosis; Neural networks; Pattern recognition
1. Introduction Many of the recent advances in spark ignition engine performance stem from, or have been made possible by, the introduction of electronic engine control (EEC) units and associated developments in automotive electronics. Early applications included the control of electronic fuel injection and spark timing. Mechanical controls have been displaced, and indeed, future legislative requirements for on-board-diagnostics and restrictions on the emission of pollutants cannot be met without electronic controls. The range of applications served by microprocessors in EEC units continues to expand, and with this, the complexity of software development and maintenance increases. This is straining conventional methods of development, and there is much interest in alternatives. The use of neural-network computing techniques is attracting * Corresponding author. Tel.: +115-951-3780; fax: +115-9513800. E-mail address: [email protected] (P.J. Shayler).
interest as one approach that may have advantages in some applications. The aim of the study reported here was to review how neural networks are being exploited, and to investigate three applications that take advantage of a range of attributes associated with these. In each application the neural networks used have a multi-layer perception (MLP) architecture and have been trained using the back-propagation algorithm. The ®rst application utilises the pattern-classi®cation abilities of a neural network in a service bay fault-diagnostic system. The second and third investigate the possibility of replacing speci®c sections of the engine control strategy software with neural-networkbased systems. The perceived advantages include generic applications of these systems to a wide range of engines, requiring only minor calibration changes for each engine modi®cation. Accordingly, the development and maintenance of the calibration and control strategy software for each new engine variant is simpler. Early examples of neural networks being exploited in automotive applications have been described by
0952-1976/00/$ - see front matter # 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 2 - 1 9 7 6 ( 9 9 ) 0 0 0 4 8 - 2
148
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Marko (1991). He developed neural-network-based systems for the real-time monitoring of vehicle control systems for faults, and the control of a non-linear active suspension system. In the second case, a backpropagation network with only two hidden nodes was trained to provide satisfactory performance using a small amount of data acquired from a vehicle travelling over `less than 100 feet of bumpy road'. Lenz and Schroeder (1996) investigated the use of the General Regression Neural Network (GRNN), described by Specht (1991), to learn certain dependencies that characterise the supply of air to an S.I. engine. These networks perform multidimensional regression between the training examples. The GRNN has two main advantages: the output is always bounded by the training examples, and it may be trained using a one-pass algorithm, greatly reducing the training time. Also it does not suer from the diculty often associated with the back-propagation algorithm, namely the convergence to a local rather than the global minimum of the error surface. The major disadvantage is that a hidden unit is required for each input pattern (unless the patterns are grouped in clusters), which can lead to prohibitively large networks if large training sets are used. Multi-layer perceptrons oer a simple structure, and may be trained by a simple algorithm in the form of back-propagation. The range of operation is limited, however, and applications are generally restricted to static mapping tasks. Beaumont and Frith (1994) proposed a solution to this problem for transient air/fuel ratio control of an S.I. engine. By applying delayed and ®ltered values of relevant measurements to its inputs, their network was provided with some knowledge of the system state or previous time history. An alternative approach to the control of dynamic plants was presented by Feldkamp et al. (1992). This addressed the problem of requiring knowledge of the recent history of inputs, outputs or internal states by using so-called `recurrent networks'. These are based on multi-layer perceptrons, but include internal feedback to units of the same or previous layers, usually with a unit time delay in the feedback loop. In this way, the network output is a function of both the current and recent inputs. The node activations provide a recent temporal history of inputs, introducing a degree of context sensitivity to the output. Two alternative schemes are used for network training: dynamic backpropagation and a technique based on the Kalman ®lter method described by Puskorius and Feldkamp (1991). The latter basically regards network training as a parameter-estimation problem, and oers superior performance and more rapid training than standard back-propagation. The two training methods were speci®cally adapted to take account of the internal state of the feedback loops. In particular, they note
that the training patterns must be presented in order, and not at random as is usually the case for conventional static mapping MLPs. Feldkamp et al. (1992) also present an optimised scheme for the development of neural controllers comprised of an identi®cation network and a controller network. The identi®cation network is initially treated as a static mapping tool, and is trained using data derived from the plant's response to a random nonrepeating input. Its purpose is to predict the state vector of the plant at the next time step, based on the current state and control input. The identi®cation network is used to calculate the gradients required for the controller training, in particular the derivative of the predicted state with respect to the trainable parameters of the controller. The controller, however, computes the control signal on the basis of the actual state, not that predicted by a model. This technique has been used with a high degree of success for many automotive-related control applications, including antilock braking systems (Davis et al., 1992), active vehicle suspensions (Puskorius and Feldkamp, 1992) and S.I. engine idle speed control (Puskorius and Feldkamp, 1993). The ®rst application presented here, in common with most documented MLP approaches, uses binary network outputs, and is concerned with pattern recognition and classi®cation. The latter two make use of analogue outputs from the networks. 2. Engine fault diagnostics 2.1. Aims This application uses the network as a pattern classi®er to assist with service-bay engine-fault diagnosis. The overall aim was to produce a system that could distinguish between intermittent and persistent ignition and fuelling faults, and indeed also identify a `normal' engine, operating correctly. Ideally, such a system should use the sensors found as standard on a production engine. The signals provided to the engine management system could then be intercepted and used by the diagnostic equipment with the minimum eort. 2.2. Data de®nition and experimental methodology The data used for diagnosis was based on the crankshaft speed ¯uctuations, over 50±100 consecutive combustion events, while the engine was idling. Several authors, including Rizzoni (1987), Mihele and Citron (1984) and Ina et al. (1986) have shown that instantaneous crankshaft speed measurements contain features related to both cylinder-speci®c and overall
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
149
Table 1 Accuracy of primary classi®cation of cylinder states for the network trained and tested on one engine (The ®nal column gives percentage of cases when either primary or secondary classi®cations were correct) Introduced fault
None (NO) Fuelling (FU) Intermittent spark (IG) Persistent mis®re (PE)
Primary classi®cation of condition (%)
PC or SC correct diagnosis (%)
NO
FU
IG
PE
99.3 61.2 0 0
0.3 30.6 12.5 0
0.3 7.1 87.5 0
0 1.2 0 100
engine performance. Shayler et al. (1990) identi®ed the dierence DS in instantaneous speed between successive minimum values as being most suitable for cylinder performance assessment under un-loaded, idle operating conditions. These minima occur at 1808 intervals of crank position, for four-cylinder engines, and DS is most dependent on the work output of the cylinder ®ring during the interval. By producing a frequency distribution of the DS values for each cylinder, Shayler et al. showed that several types of cylinderspeci®c faults could be identi®ed. This method was found to be robust when tested on a variety of four cylinder engine types with capacities ranging from 1.1 to 2.0 l. Fig. 1. shows typical DS frequency distributions obtained under four-cylinder conditions: fault-free, intermittent mis®re due to incorrect cylinder air/fuel ratio (AFR), intermittent mis®re due to ignition defect, and persistent mis®re. Each subdivision of the DS range represents an interval of 10 rev/min. The data were acquired from tests on a ¯eet of thirty vehicles. Although there are clear dierences between the mean distribution shapes associated with each cylinder condition, accurately identifying this condition in a particular case is made dicult by the range of values that each of the histogram bars can take on. The network used for this classi®cation problem has a 7:7:4 con®guration, where the numbers refer to the number of input units, hidden layer units, and output units respectively. Each input unit was associated with one range of the DS histogram. Each output unit was associated with one type of cylinder condition or fault. Two networks were trained separately on dierent sets of input/output pattern pairs. The ®rst set con-
99.3 70.6 87.5 100
tained examples from only one engine. Eight example patterns for each cylinder condition were used to train the network. The remaining patterns were reserved to test the classi®cation accuracy of the trained network. The most probable cylinder state was associated with the output unit with the highest activation values. This is termed the `primary classi®cation'. In instances where the wrong diagnosis was indicated but the second highest output was signi®cant and correct, this was logged as the correct secondary classi®cation. The results of this assessment of diagnostic performance is summarised in Table 1. Almost all of the fault-free (NO) and persistent mis®re (PE) cylinder conditions were correctly identi®ed. Of the ignition mis®re (IG) cases, 87.5% were identi®ed, with the remainder associated with fuelling faults (FU). Overall, the network gave a primary classi®cation accuracy (PCA) of 79.4% and a secondary classi®cation accuracy (SCA) of 10.0%. Only 10.7% of conditions were not correctly identi®ed by either the primary or secondary classi®cation. The second network was trained on examples acquired from twelve engines of the same basic type but with capacities of 1.6, 1.8 and 2.0 l. The aim of this work was to assess whether the classi®cation accuracy observed in the case above, for a single engine, could be maintained across a range of engines. Tests were carried out on each engine, with and without faults introduced, to generate a database of more than 3000 DS histogram patterns. The network was trained on thirty example patterns for each cylinder condition. The remaining patterns were then used in the assessment of network performance. This is summarised in Table 2.
Table 2 Accuracy of primary classi®cation of cylinder states for the network trained and tested on twelve engines Introduced fault
None (NO) Fuelling (FU) Intermittent spark (IG) Persistent mis®re (PE)
Primary classi®cation of condition (%)
PC or SC correct diagnosis (%)
NO
FU
IG
PE
88.2 24.9 3.5 0
9.9 59.8 12.8 0.8
1.9 13.0 82.9 0.8
0 2.4 0.8 98.4
96.8 88.2 87.2 99.2
150
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Fig. 1. DS distribution patterns associated with cylinder-speci®c faults. Shaded bars show mean distributions. Open bars show minimum and maximum values in data sets. Cylinder states are as follows: A, Fault-free. B, Occasional mis®re due to fuelling fault. C, Frequent mis®re due to fuelling fault. D, Intermittent mis®re due to ignition fault. E, Persistent mis®re (various causes).
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
2.3. Summary of performance The results summarised in Table 2 demonstrate what can be achieved using a relatively simple, 7:7:4 network. The variability of patterns associated with particularly cylinder conditions was representative of engine-to-engine dierences due to production tolerances, accumulated mileage, and wear. The overall primary classi®cation accuracy was 82.3%, compared to 79.4% for the network trained and tested on data from only one engine. The improvement is associated with a higher success rate in diagnosing fuelling faults. Work carried out on an even wider range of fourcylinder engine types has shown that similar levels of performance can be achieved. This range included engines with manual, conventional automatic, or continuously variable transmissions. 3. Steady-state control 3.1. Background This application was primarily concerned with validating the accuracy of neural networks applied to mapping problems encountered in EECs. The approach taken was to train neural networks to emulate elements of a current production engine control strategy. Attention was focused on the control of steady-state part-load operation with feedforward fuelling control. Data recorded during rolling road tests were used to train the networks, and also to provide a standard against which the accuracy of the trained network could be measured. The control of the fuel injector pulsewidth and ignition spark timing was identi®ed as the target that networks might be trained to perform to some advantage. By conventional methods this requires the greatest mapping and calibration eort in current EEC systems. Analysis of the EEC-IV strategy revealed that only four key input signals were driving the process by which the pulse width and spark timing were set. Thus, a single network with four inputs and two outputs could replace several lookup tables and calibration functions used in the EEC-IV system. Three types of network implementation were examined, all of which used the same basic, four-input, twooutput mapping. A single hidden layer of four units was also used in all three cases. The ®rst used the 4:4:2 network to provide outputs that were directly proportional to the engineering unit values. Thus, calculation of the outputs in the required units is simply a matter of rescaling the network output values. Both alternative methods used simple mathematical functions to produce approximations to the required outputs. The basis for this approach was the assumption
151
that standard approximation functions could be developed for a wide range of engines, with the strategy being tuned for particular engine variants using neural networks trained to give appropriate correction values. One method simply added the network output to the approximate function; the other made corrections in proportion to the approximate value. The output scaling operations used with each of these methods are described in more detail in another publication (Bacon et al., 1992). The four input signals to the network were derived from sensed values (air charge temperature, air mass ¯ow rate, engine speed, and normalised air charge). The data sets of inputs and outputs (fuel injector pulsewidth and spark timing) used to train and test the network were generated using a 1.8 l high-output Zetec engine in a vehicle driven through a range of load and speed conditions. Data from the EEC system were logged at 19 test conditions, each providing around 30 input/output sets of training examples. The test conditions covered engine speeds ranging from 1000 to 4000 rev/min, and intake manifold air pressures between 0.35 and 0.7 bar. These represent a range of light-load to medium-load operating conditions. A further ®ve tests were used to generate performance veri®cation data. The engine speed and intake pressure settings for these test points were respectively: (I) 1500 rev/min, 0.4 bar; (II) 1500 rev/min, 0.65 bar; (III) 2500 rev/min, 0.55 bar; (IV) 2500 rev/min, 0.75 bar; and (V) 3500 rev/min, 0.55 bar. 3.2. Summary of performance After training on data from the ®rst set of test conditions, the various systems were tested on their ability to produce the required output at a range of dierent operation conditions. Figs. 2 and 3 compare the spark advance and fuel pulsewidth results obtained from the trained neural network and EEC-IV at ®ve test conditions. The overall RMS errors are 34.1% for spark advance, and 5.4% for fuelling. The overall error for fuelling is largely due to that arising at condition IV (11.4%) where the network must extrapolate from, rather than interpolate between, the test data. The larger errors in spark timing are attributable to the greater complexity of the target spark timing map used in the EEC system. Increasing the density of the training data to reduce errors was an unattractive option, since this requires more engine testing to calibrate each new engine variant. Hence, the use of approximate functions corrected by the network was examined. Figs. 4 and 5 show the results of testing the addition correction network. The total output (function + network) is compared with the desired values. Again, the fuel pulsewidth values are
152
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Fig. 2. Absolute value 4:4:2 network: accuracy of spark advance output.
Fig. 3. Absolute value 4:4:2 network: accuracy of fuel pulsewidth output.
Fig. 4. Addition correction 4:4:2 network: accuracy of spark advance output.
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
153
Fig. 5. Addition correction 4:4:2 network: accuracy of fuel pulsewidth output.
more accurate, having an RMS error of 2.3% compared with 17.9% for the spark timing output. The proportional correction approach prevented large corrections being made to small values, and so reduced the RMS errors to 11.8% for spark advance, and 1.5% for fuelling (see Figs. 6 and 7). Although the system does not perfectly reproduce the EEC-IV response, the errors are reduced to a level that should allow an engine to be controlled by the system. The results indicate that neural-network systems can be used to replace the combination of control algorithms and look-up tables found in a conventional EEC system. In this way the development eort and expertise required to calibrate the control system for a new engine variant is greatly reduced.
4. Transient engine control 4.1. Background The multi-point fuel injection system described above performs very well when the engine is operating under steady-state conditions. Unfortunately problems arise when a change in fuel delivery rate is required. As indicated in Fig. 8, a signi®cant proportion of the injected fuel is deposited on the intake port surface, causing a delay in the transport of this fuel to the combustion chamber. Consequently, changes in the rate of fuel injection do not immediately produce an equal change in the inducted fuel ¯ow rate. To maintain a constant air/fuel ratio while the air ¯ow rate is changing, the injected fuel ¯ow rate must be adjusted
Fig. 6. Proportional correction 4:4:2 network: accuracy of spark advance output.
154
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Fig. 7. Proportional correction 4:4:2 network: accuracy of fuel pulsewidth output.
to account for the eect of the fuel ®lms on the port surface. The change in fuel injected to provide this adjustment is referred to as transient fuel compensation (TFC). A commonly adopted approach to de®ning the TFC requirements is based on a phenomenological model of fuel transport in the intake port, originally proposed by Hires and Overington (1981). Developments and applications of this have been reported by Aquino (1981) and by Shayler et al. (1995). The model represents the dynamics of the fuel transfer from injector to cylinder as governed by two equations involving two empirical parameters t and X. The time constant t is associated with the rate of fuel transfer from surface ®lms, and X is the fraction of fuel injected which is deposited in the ®lms. The governing equations are: m_ ind 1 ÿ X m_ inj
mff t
1
Fig. 8. Schematic diagram of the inlet port and the wall t ÿ X wetting model, illustrating the formation of fuel-®lms on both intake valve and port.
and m_ ff Xm_ inj ÿ
mff t
2
where m_ ind is the rate of fuel induction, m_ inj is the rate of fuel injection and mff is the fuel ®lm mass. From Eq. (1) the instantaneous requirement for TFC can be deduced as: m_ tfc
1 mff Xm_ base ÿ 1 ÿ X t
3
where m_ tfc is the compensation fuel ¯ow rate added to the base fuel ¯ow rate m_ base : The base fuel is that required in the absence of fuel transport eects. The application of this model in an EEC system entails considerable calibration work initially, and signi®cant onboard computing eort in service. The attractions of considering the use of neural networks for de®ning and calibrating a TFC scheme are the simpli®cations to the development process and implementation that might be achieved, since a neural network is able to learn the output that should be provided for a particular input condition, given that sucient training examples can be presented to describe that relationship. The work presented here demonstrates how a neural network can be used to calculate the TFC fuel required for the prevailing input conditions, such that only one pass through the network is necessary each time a new output is required. This contrasts with the method of Beaumont and Frith (1994), which used a neural network as a model of the engine, and required an iterative procedure to calculate the required fuelling for the prevailing operating conditions. Calibration eort is still required to establish a database for network training, covering a range of input conditions and the associated TFC requirements, but an experimental methodology has been developed
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
that assumes only a minimum knowledge of the ®lm dynamics. Study of control strategy requirements and of the expression for t and X referred to above suggests that the required TFC is a function of three main parameters. These are normalised air ¯ow rate, engine coolant temperature and fuel ®lm mass. The air ¯ow rate was normalised with respect to the maximum that would occur if the engine ran at the same speed with the cylinders ®lled with air at standard temperature and pressure. Coolant temperature was chosen as being representative of intake manifold temperature. It is not an accurate measure, but is available to the existing control strategy. The fuel ®lm mass was estimated by integration of the TFC fuel output, thus introducing state feedback to deal with the dynamic nature of the problem. As a result, TFC is applied until the calculated fuel ®lm mass reaches the value for which the network has learned to supply zero TFC fuel for that air ¯ow and temperature. Since TFC fuel decreases with increasing fuel ®lm mass, the supplied TFC will always converge to zero under steady-state operating conditions, even if the network output is subject to some error. Clearly, fuel ¯ow per unit time is dependent upon speed, but fuel mass per injection for a given cylinder air charge may be independent of speed. Thus, speed has not been used as an input to the neural network at present. Future work may, however, show that it is required to produce the required AFR accuracy. 4.2. Experimental methodology For engine test purposes the TFC neural networks were trained according to a procedure that avoids prior knowledge of intake port fuel transport characteristics. This method is described in detail in (Shayler et al., 1996), but may be summarised as follows. Only a basic understanding of the physical behaviour of the system is required, but as a result several approximations must be made. These necessitate the adoption of an iterative approach to determine the ®nal, correct TFC requirements. It was assumed that, since the TFC fuel is supplied to compensate for changes in the fuel ®lm mass, for a correctly compensated transient, the total change in fuel ®lm mass will be equal to the TFC fuel supplied. Also, changes to the mass of fuel entering or leaving the fuel ®lm at any given point in the transient, as a result of changing the injected fuel, are ignored. The latter of these two will give an inaccuracy in the calculated TFC requirement, but this will be less than that present in the TFC applied in the preceding test, and an iterative procedure should converge to the actual TFC requirement. By making these assumptions, the correction to the TFC requirement simpli®es to the dierence between the mass of fuel required to
155
give a stoichiometric mixture and the actual mass of fuel entering the combustion cylinder during the previous test. This actual mass was determined from the known cylinder air charge and the corresponding exhaust AFR. From this new TFC requirement, the corresponding fuel ®lm mass variation required as a network input was determined by integration with respect to injection number. The initial fuel ®lm mass was calculated from the relationship between the overall change in fuel ®lm mass and the change in air ¯ow rate for the training transients. Since three transient pairs were used, a quadratic relationship was required, with the assumption that the fuel ®lm mass was zero for zero air ¯ow. A simple plenum ®lling/emptying model of the inlet manifold, together with electronic throttle control, was used to predict air ¯ow at the time of fuel induction. Such a scheme is necessary to allow time for the fuel requirement to be calculated, injected and transported to the cylinder, for induction in phase with air ¯ow. 4.3. Summary of performance To facilitate the development and evaluation of the new fuelling control strategy, control of the fuel injectors was taken over by software running on an IBMcompatible 486 personal computer (PC). An exhaust AFR sensor was also installed to assess the transient fuel compensation accuracy. Electronic throttle control was provided by a positioning stepper motor driven by the fuelling software. The performance of the neuralnetwork-based TFC scheme was assessed for a range of transients. The experimental work was carried out on a 1.8 l Ford Zetec engine. The results of these tests are illustrated in Fig. 9, in terms of peak AFR excursion. The results include data for engine speeds of 1000, 1500 and 2000 rpm, for a network trained at only 1000 rpm. At each speed, the tests covered a range of air ¯ow changes both within and extending beyond the limits used to generate training data. From this ®gure it can be seen that the peak excursion is below 1.2 units in the majority of cases, there being only two transients for which the excursion is larger, but still less than 1.8 units, and these only occur when the lower normalised air ¯ow rate of the transient is below 0.25. These results compare well with those for the present production control system for the same engine, which gives typical excursions of up to 2.4 units with some as large as 4 units. The network-based system has a simpler structure and eliminates a number of look-up tables found in a more conventional system. There is little calculation required in addition to the propagation of data through the trained network. The relevant inputs (normalised air ¯ow, coolant temperature, and estimated
156
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
found to be more successful. Addition and proportional correction schemes similar to those described for the previous application were examined (Goodman, 1996), but were found to be less successful, due to the increased complexity of the relationship between network input and output values.
5. Discussion and conclusions The introduction of microprocessor-based, electronic engine control units has enabled improvements to be made in many areas of systems control and monitoring. New applications and further improvements continue to emerge. The development of control strategy and, more generally, engine management software is an expanding task. This may contain literally hundreds of parameters which must be calibrated for each application. The expertise required is considerable, as are the resources vested in documenting and supporting applications. Set against this background, neural networks have attracted interest for reasons that include the possible simpli®cation of software development and calibration, and the utility of their pattern-classi®cation characteristics. As various investigators have reported, these have been exploited in a number of engine and vehicle systems applications. In the work reported here, three engine-related applications have been examined. These illustrate the diversity of problems and ways in which neural networks can be exploited. The results obtained using relatively simple networks and network training procedures have been generally encouraging. As important as the expertise in the neural-network ®eld, has been an understanding of the nature of the problem being addressed. The likelihood of a successful implementation depends on both.
Acknowledgements
Fig. 9. Summary of performance of neural-network-based compensation over a range of air ¯ow changes and engine speeds. The network was only trained on data from 1000 rpm tests.
fuel ®lm mass) require linear scaling from real units to valid network units, the converse being true for the TFC fuel mass output, the only other necessary calculation being to modify the estimate of fuel ®lm mass for the cylinder concerned by the addition of the calculated TFC fuel mass. For this application an absolute value network was
The authors wish to acknowledge and thank the Ford Motor Company for ®nancial support and for permission to publish this work.
References Aquino, C.F., 1981. Transient A/F control characteristics of the 5 l central fuel injection engine. SAE Technical Paper No. 810494. Bacon, A., Shayler, P.J., Ma, T., 1992. Potential for engine control using neural networks. IMechE Paper No. C448/057. Beaumont, A.J., Frith, A.M., 1994. Adaptive control of transient air±fuel ratio using neural networks. In: Proceedings of the 27th International Symposium on Automotive Technology and Automation (ISATA), Aachen, Germany. Davis Jr, L.I., Puskorius, G.V., Yuan, F., Feldkamp, L.A., 1992.
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157 Neural network modelling and control of an anti-lock brake system. In: Proceedings of Intelligent Vehicles '92 Symposium, Detroit. Feldkamp, L.A., Puskorius, G.V., Davis, L.I., Yuan, F., 1992. Neural control systems trained by dynamic gradient methods for automotive applications. International Joint Conference on Neural Networks 2, 798±804. Goodman, M.S., 1996. Transient air/fuel ratio control of spark ignition engines using neural networks. Ph.D. thesis, University of Nottingham. Hires, S.D., Overington, M.T., 1981. Transient mixture strength excursions Ð an investigation of their causes and the development of a constant mixture strength fuelling strategy. SAE Technical Paper No. 810495. Ina, T., Obayahsi, H., Shigematsu, T., 1986. Lean limit AFR control using engine speed variation. SAE Technical Paper No. 860413. Lenz, U., Schroeder, D., 1996. Arti®cial intelligence for combustion engine control. SAE Technical Paper No. 960328. Marko, K.A., 1991. Neural network application to diagnostics and control of vehicle control systems. In: Neural Information Processing Symposium (NIPS-91). Morgan Kaufman, New York. Mihele, W.P., Citron, S.J., 1984. An on-line roughness measurements technique. SAE Technical Paper No. 840316. Puskorius, G.V., Feldkamp, L.A., 1991. Decoupled extended
157
Kalman ®lter training of feedforward layered networks. In: International Joint Conference on Neural Networks, vol. I, IEEE, New York, 771±777. Puskorius, G.V., Feldkamp, L.A., 1992. Model reference adaptive control with recurrent networks trained by the dynamic DEKF algorithm. In: Proceedings of the 1992 International Joint Conference on Neural Networks, Baltimore. Puskorius, G.V., Feldkamp, L.A., 1993. Automotive engine idle speed control with recurrent neural networks. In: Proceedings of the 1993 American Control Conference, San Francisco. Rizzoni, F., 1987. Test transforms for rapid isolation of mis®ring cylinder. SAE Technical Paper No. 871915. Shayler, P.J., White, D.M., Wilkinson, A.C., Wakeling, R. 1990. The assessment of information sources for fault detection on spark ignition engines. IMechE Paper No. C422/006. Shayler, P.J., Goodman, M.S., Ma, T., 1996. Transient air/fuel ratio control of an S.I. engine using neural networks. SAE Technical Paper No. 960326. Shayler, P.J., Teo, Y.C., Scarisbrick, A., 1995. Fuel transport characteristics of spark ignition engines for transient fuel compensation. SAE Technical Paper No. 950067. Specht, D.F., 1991. A general regression neural network. IEEE Transactions on Neural Networks 2 (6), 568±576.
www.elsevier.com/locate/engappai
The exploitation of neural networks in automotive engine management systems P.J. Shayler a,*, M. Goodman a, T. Ma b a
University of Nottingham, Nottingham, NG7 2RD, UK Ford Motor Company Research and Engineering Centre, Essex, SS15 6EE, UK
b
Received 1 October 1998; accepted 1 October 1999
Abstract The use of electronic engine control systems on spark ignition engines has enabled a high degree of performance optimisation to be achieved. The range of functions performed by these systems, and the level of performance demanded, is rising and thus so are development times and costs. Neural networks have attracted attention as having the potential to simplify software development and improve the performance of this software. The scope and nature of possible applications is described. In particular, the pattern recognition and classi®cation abilities of networks are applied to crankshaft speed ¯uctuation data for engine-fault diagnosis, and multidimensional mapping capabilities are investigated as an alternative to large `lookup' tables and calibration functions. # 2000 Elsevier Science Ltd. All rights reserved. Keywords: Automotive engines; Calibration; Engine control; Fault diagnosis; Neural networks; Pattern recognition
1. Introduction Many of the recent advances in spark ignition engine performance stem from, or have been made possible by, the introduction of electronic engine control (EEC) units and associated developments in automotive electronics. Early applications included the control of electronic fuel injection and spark timing. Mechanical controls have been displaced, and indeed, future legislative requirements for on-board-diagnostics and restrictions on the emission of pollutants cannot be met without electronic controls. The range of applications served by microprocessors in EEC units continues to expand, and with this, the complexity of software development and maintenance increases. This is straining conventional methods of development, and there is much interest in alternatives. The use of neural-network computing techniques is attracting * Corresponding author. Tel.: +115-951-3780; fax: +115-9513800. E-mail address: [email protected] (P.J. Shayler).
interest as one approach that may have advantages in some applications. The aim of the study reported here was to review how neural networks are being exploited, and to investigate three applications that take advantage of a range of attributes associated with these. In each application the neural networks used have a multi-layer perception (MLP) architecture and have been trained using the back-propagation algorithm. The ®rst application utilises the pattern-classi®cation abilities of a neural network in a service bay fault-diagnostic system. The second and third investigate the possibility of replacing speci®c sections of the engine control strategy software with neural-networkbased systems. The perceived advantages include generic applications of these systems to a wide range of engines, requiring only minor calibration changes for each engine modi®cation. Accordingly, the development and maintenance of the calibration and control strategy software for each new engine variant is simpler. Early examples of neural networks being exploited in automotive applications have been described by
0952-1976/00/$ - see front matter # 2000 Elsevier Science Ltd. All rights reserved. PII: S 0 9 5 2 - 1 9 7 6 ( 9 9 ) 0 0 0 4 8 - 2
148
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Marko (1991). He developed neural-network-based systems for the real-time monitoring of vehicle control systems for faults, and the control of a non-linear active suspension system. In the second case, a backpropagation network with only two hidden nodes was trained to provide satisfactory performance using a small amount of data acquired from a vehicle travelling over `less than 100 feet of bumpy road'. Lenz and Schroeder (1996) investigated the use of the General Regression Neural Network (GRNN), described by Specht (1991), to learn certain dependencies that characterise the supply of air to an S.I. engine. These networks perform multidimensional regression between the training examples. The GRNN has two main advantages: the output is always bounded by the training examples, and it may be trained using a one-pass algorithm, greatly reducing the training time. Also it does not suer from the diculty often associated with the back-propagation algorithm, namely the convergence to a local rather than the global minimum of the error surface. The major disadvantage is that a hidden unit is required for each input pattern (unless the patterns are grouped in clusters), which can lead to prohibitively large networks if large training sets are used. Multi-layer perceptrons oer a simple structure, and may be trained by a simple algorithm in the form of back-propagation. The range of operation is limited, however, and applications are generally restricted to static mapping tasks. Beaumont and Frith (1994) proposed a solution to this problem for transient air/fuel ratio control of an S.I. engine. By applying delayed and ®ltered values of relevant measurements to its inputs, their network was provided with some knowledge of the system state or previous time history. An alternative approach to the control of dynamic plants was presented by Feldkamp et al. (1992). This addressed the problem of requiring knowledge of the recent history of inputs, outputs or internal states by using so-called `recurrent networks'. These are based on multi-layer perceptrons, but include internal feedback to units of the same or previous layers, usually with a unit time delay in the feedback loop. In this way, the network output is a function of both the current and recent inputs. The node activations provide a recent temporal history of inputs, introducing a degree of context sensitivity to the output. Two alternative schemes are used for network training: dynamic backpropagation and a technique based on the Kalman ®lter method described by Puskorius and Feldkamp (1991). The latter basically regards network training as a parameter-estimation problem, and oers superior performance and more rapid training than standard back-propagation. The two training methods were speci®cally adapted to take account of the internal state of the feedback loops. In particular, they note
that the training patterns must be presented in order, and not at random as is usually the case for conventional static mapping MLPs. Feldkamp et al. (1992) also present an optimised scheme for the development of neural controllers comprised of an identi®cation network and a controller network. The identi®cation network is initially treated as a static mapping tool, and is trained using data derived from the plant's response to a random nonrepeating input. Its purpose is to predict the state vector of the plant at the next time step, based on the current state and control input. The identi®cation network is used to calculate the gradients required for the controller training, in particular the derivative of the predicted state with respect to the trainable parameters of the controller. The controller, however, computes the control signal on the basis of the actual state, not that predicted by a model. This technique has been used with a high degree of success for many automotive-related control applications, including antilock braking systems (Davis et al., 1992), active vehicle suspensions (Puskorius and Feldkamp, 1992) and S.I. engine idle speed control (Puskorius and Feldkamp, 1993). The ®rst application presented here, in common with most documented MLP approaches, uses binary network outputs, and is concerned with pattern recognition and classi®cation. The latter two make use of analogue outputs from the networks. 2. Engine fault diagnostics 2.1. Aims This application uses the network as a pattern classi®er to assist with service-bay engine-fault diagnosis. The overall aim was to produce a system that could distinguish between intermittent and persistent ignition and fuelling faults, and indeed also identify a `normal' engine, operating correctly. Ideally, such a system should use the sensors found as standard on a production engine. The signals provided to the engine management system could then be intercepted and used by the diagnostic equipment with the minimum eort. 2.2. Data de®nition and experimental methodology The data used for diagnosis was based on the crankshaft speed ¯uctuations, over 50±100 consecutive combustion events, while the engine was idling. Several authors, including Rizzoni (1987), Mihele and Citron (1984) and Ina et al. (1986) have shown that instantaneous crankshaft speed measurements contain features related to both cylinder-speci®c and overall
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
149
Table 1 Accuracy of primary classi®cation of cylinder states for the network trained and tested on one engine (The ®nal column gives percentage of cases when either primary or secondary classi®cations were correct) Introduced fault
None (NO) Fuelling (FU) Intermittent spark (IG) Persistent mis®re (PE)
Primary classi®cation of condition (%)
PC or SC correct diagnosis (%)
NO
FU
IG
PE
99.3 61.2 0 0
0.3 30.6 12.5 0
0.3 7.1 87.5 0
0 1.2 0 100
engine performance. Shayler et al. (1990) identi®ed the dierence DS in instantaneous speed between successive minimum values as being most suitable for cylinder performance assessment under un-loaded, idle operating conditions. These minima occur at 1808 intervals of crank position, for four-cylinder engines, and DS is most dependent on the work output of the cylinder ®ring during the interval. By producing a frequency distribution of the DS values for each cylinder, Shayler et al. showed that several types of cylinderspeci®c faults could be identi®ed. This method was found to be robust when tested on a variety of four cylinder engine types with capacities ranging from 1.1 to 2.0 l. Fig. 1. shows typical DS frequency distributions obtained under four-cylinder conditions: fault-free, intermittent mis®re due to incorrect cylinder air/fuel ratio (AFR), intermittent mis®re due to ignition defect, and persistent mis®re. Each subdivision of the DS range represents an interval of 10 rev/min. The data were acquired from tests on a ¯eet of thirty vehicles. Although there are clear dierences between the mean distribution shapes associated with each cylinder condition, accurately identifying this condition in a particular case is made dicult by the range of values that each of the histogram bars can take on. The network used for this classi®cation problem has a 7:7:4 con®guration, where the numbers refer to the number of input units, hidden layer units, and output units respectively. Each input unit was associated with one range of the DS histogram. Each output unit was associated with one type of cylinder condition or fault. Two networks were trained separately on dierent sets of input/output pattern pairs. The ®rst set con-
99.3 70.6 87.5 100
tained examples from only one engine. Eight example patterns for each cylinder condition were used to train the network. The remaining patterns were reserved to test the classi®cation accuracy of the trained network. The most probable cylinder state was associated with the output unit with the highest activation values. This is termed the `primary classi®cation'. In instances where the wrong diagnosis was indicated but the second highest output was signi®cant and correct, this was logged as the correct secondary classi®cation. The results of this assessment of diagnostic performance is summarised in Table 1. Almost all of the fault-free (NO) and persistent mis®re (PE) cylinder conditions were correctly identi®ed. Of the ignition mis®re (IG) cases, 87.5% were identi®ed, with the remainder associated with fuelling faults (FU). Overall, the network gave a primary classi®cation accuracy (PCA) of 79.4% and a secondary classi®cation accuracy (SCA) of 10.0%. Only 10.7% of conditions were not correctly identi®ed by either the primary or secondary classi®cation. The second network was trained on examples acquired from twelve engines of the same basic type but with capacities of 1.6, 1.8 and 2.0 l. The aim of this work was to assess whether the classi®cation accuracy observed in the case above, for a single engine, could be maintained across a range of engines. Tests were carried out on each engine, with and without faults introduced, to generate a database of more than 3000 DS histogram patterns. The network was trained on thirty example patterns for each cylinder condition. The remaining patterns were then used in the assessment of network performance. This is summarised in Table 2.
Table 2 Accuracy of primary classi®cation of cylinder states for the network trained and tested on twelve engines Introduced fault
None (NO) Fuelling (FU) Intermittent spark (IG) Persistent mis®re (PE)
Primary classi®cation of condition (%)
PC or SC correct diagnosis (%)
NO
FU
IG
PE
88.2 24.9 3.5 0
9.9 59.8 12.8 0.8
1.9 13.0 82.9 0.8
0 2.4 0.8 98.4
96.8 88.2 87.2 99.2
150
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Fig. 1. DS distribution patterns associated with cylinder-speci®c faults. Shaded bars show mean distributions. Open bars show minimum and maximum values in data sets. Cylinder states are as follows: A, Fault-free. B, Occasional mis®re due to fuelling fault. C, Frequent mis®re due to fuelling fault. D, Intermittent mis®re due to ignition fault. E, Persistent mis®re (various causes).
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
2.3. Summary of performance The results summarised in Table 2 demonstrate what can be achieved using a relatively simple, 7:7:4 network. The variability of patterns associated with particularly cylinder conditions was representative of engine-to-engine dierences due to production tolerances, accumulated mileage, and wear. The overall primary classi®cation accuracy was 82.3%, compared to 79.4% for the network trained and tested on data from only one engine. The improvement is associated with a higher success rate in diagnosing fuelling faults. Work carried out on an even wider range of fourcylinder engine types has shown that similar levels of performance can be achieved. This range included engines with manual, conventional automatic, or continuously variable transmissions. 3. Steady-state control 3.1. Background This application was primarily concerned with validating the accuracy of neural networks applied to mapping problems encountered in EECs. The approach taken was to train neural networks to emulate elements of a current production engine control strategy. Attention was focused on the control of steady-state part-load operation with feedforward fuelling control. Data recorded during rolling road tests were used to train the networks, and also to provide a standard against which the accuracy of the trained network could be measured. The control of the fuel injector pulsewidth and ignition spark timing was identi®ed as the target that networks might be trained to perform to some advantage. By conventional methods this requires the greatest mapping and calibration eort in current EEC systems. Analysis of the EEC-IV strategy revealed that only four key input signals were driving the process by which the pulse width and spark timing were set. Thus, a single network with four inputs and two outputs could replace several lookup tables and calibration functions used in the EEC-IV system. Three types of network implementation were examined, all of which used the same basic, four-input, twooutput mapping. A single hidden layer of four units was also used in all three cases. The ®rst used the 4:4:2 network to provide outputs that were directly proportional to the engineering unit values. Thus, calculation of the outputs in the required units is simply a matter of rescaling the network output values. Both alternative methods used simple mathematical functions to produce approximations to the required outputs. The basis for this approach was the assumption
151
that standard approximation functions could be developed for a wide range of engines, with the strategy being tuned for particular engine variants using neural networks trained to give appropriate correction values. One method simply added the network output to the approximate function; the other made corrections in proportion to the approximate value. The output scaling operations used with each of these methods are described in more detail in another publication (Bacon et al., 1992). The four input signals to the network were derived from sensed values (air charge temperature, air mass ¯ow rate, engine speed, and normalised air charge). The data sets of inputs and outputs (fuel injector pulsewidth and spark timing) used to train and test the network were generated using a 1.8 l high-output Zetec engine in a vehicle driven through a range of load and speed conditions. Data from the EEC system were logged at 19 test conditions, each providing around 30 input/output sets of training examples. The test conditions covered engine speeds ranging from 1000 to 4000 rev/min, and intake manifold air pressures between 0.35 and 0.7 bar. These represent a range of light-load to medium-load operating conditions. A further ®ve tests were used to generate performance veri®cation data. The engine speed and intake pressure settings for these test points were respectively: (I) 1500 rev/min, 0.4 bar; (II) 1500 rev/min, 0.65 bar; (III) 2500 rev/min, 0.55 bar; (IV) 2500 rev/min, 0.75 bar; and (V) 3500 rev/min, 0.55 bar. 3.2. Summary of performance After training on data from the ®rst set of test conditions, the various systems were tested on their ability to produce the required output at a range of dierent operation conditions. Figs. 2 and 3 compare the spark advance and fuel pulsewidth results obtained from the trained neural network and EEC-IV at ®ve test conditions. The overall RMS errors are 34.1% for spark advance, and 5.4% for fuelling. The overall error for fuelling is largely due to that arising at condition IV (11.4%) where the network must extrapolate from, rather than interpolate between, the test data. The larger errors in spark timing are attributable to the greater complexity of the target spark timing map used in the EEC system. Increasing the density of the training data to reduce errors was an unattractive option, since this requires more engine testing to calibrate each new engine variant. Hence, the use of approximate functions corrected by the network was examined. Figs. 4 and 5 show the results of testing the addition correction network. The total output (function + network) is compared with the desired values. Again, the fuel pulsewidth values are
152
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Fig. 2. Absolute value 4:4:2 network: accuracy of spark advance output.
Fig. 3. Absolute value 4:4:2 network: accuracy of fuel pulsewidth output.
Fig. 4. Addition correction 4:4:2 network: accuracy of spark advance output.
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
153
Fig. 5. Addition correction 4:4:2 network: accuracy of fuel pulsewidth output.
more accurate, having an RMS error of 2.3% compared with 17.9% for the spark timing output. The proportional correction approach prevented large corrections being made to small values, and so reduced the RMS errors to 11.8% for spark advance, and 1.5% for fuelling (see Figs. 6 and 7). Although the system does not perfectly reproduce the EEC-IV response, the errors are reduced to a level that should allow an engine to be controlled by the system. The results indicate that neural-network systems can be used to replace the combination of control algorithms and look-up tables found in a conventional EEC system. In this way the development eort and expertise required to calibrate the control system for a new engine variant is greatly reduced.
4. Transient engine control 4.1. Background The multi-point fuel injection system described above performs very well when the engine is operating under steady-state conditions. Unfortunately problems arise when a change in fuel delivery rate is required. As indicated in Fig. 8, a signi®cant proportion of the injected fuel is deposited on the intake port surface, causing a delay in the transport of this fuel to the combustion chamber. Consequently, changes in the rate of fuel injection do not immediately produce an equal change in the inducted fuel ¯ow rate. To maintain a constant air/fuel ratio while the air ¯ow rate is changing, the injected fuel ¯ow rate must be adjusted
Fig. 6. Proportional correction 4:4:2 network: accuracy of spark advance output.
154
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
Fig. 7. Proportional correction 4:4:2 network: accuracy of fuel pulsewidth output.
to account for the eect of the fuel ®lms on the port surface. The change in fuel injected to provide this adjustment is referred to as transient fuel compensation (TFC). A commonly adopted approach to de®ning the TFC requirements is based on a phenomenological model of fuel transport in the intake port, originally proposed by Hires and Overington (1981). Developments and applications of this have been reported by Aquino (1981) and by Shayler et al. (1995). The model represents the dynamics of the fuel transfer from injector to cylinder as governed by two equations involving two empirical parameters t and X. The time constant t is associated with the rate of fuel transfer from surface ®lms, and X is the fraction of fuel injected which is deposited in the ®lms. The governing equations are: m_ ind 1 ÿ X m_ inj
mff t
1
Fig. 8. Schematic diagram of the inlet port and the wall t ÿ X wetting model, illustrating the formation of fuel-®lms on both intake valve and port.
and m_ ff Xm_ inj ÿ
mff t
2
where m_ ind is the rate of fuel induction, m_ inj is the rate of fuel injection and mff is the fuel ®lm mass. From Eq. (1) the instantaneous requirement for TFC can be deduced as: m_ tfc
1 mff Xm_ base ÿ 1 ÿ X t
3
where m_ tfc is the compensation fuel ¯ow rate added to the base fuel ¯ow rate m_ base : The base fuel is that required in the absence of fuel transport eects. The application of this model in an EEC system entails considerable calibration work initially, and signi®cant onboard computing eort in service. The attractions of considering the use of neural networks for de®ning and calibrating a TFC scheme are the simpli®cations to the development process and implementation that might be achieved, since a neural network is able to learn the output that should be provided for a particular input condition, given that sucient training examples can be presented to describe that relationship. The work presented here demonstrates how a neural network can be used to calculate the TFC fuel required for the prevailing input conditions, such that only one pass through the network is necessary each time a new output is required. This contrasts with the method of Beaumont and Frith (1994), which used a neural network as a model of the engine, and required an iterative procedure to calculate the required fuelling for the prevailing operating conditions. Calibration eort is still required to establish a database for network training, covering a range of input conditions and the associated TFC requirements, but an experimental methodology has been developed
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
that assumes only a minimum knowledge of the ®lm dynamics. Study of control strategy requirements and of the expression for t and X referred to above suggests that the required TFC is a function of three main parameters. These are normalised air ¯ow rate, engine coolant temperature and fuel ®lm mass. The air ¯ow rate was normalised with respect to the maximum that would occur if the engine ran at the same speed with the cylinders ®lled with air at standard temperature and pressure. Coolant temperature was chosen as being representative of intake manifold temperature. It is not an accurate measure, but is available to the existing control strategy. The fuel ®lm mass was estimated by integration of the TFC fuel output, thus introducing state feedback to deal with the dynamic nature of the problem. As a result, TFC is applied until the calculated fuel ®lm mass reaches the value for which the network has learned to supply zero TFC fuel for that air ¯ow and temperature. Since TFC fuel decreases with increasing fuel ®lm mass, the supplied TFC will always converge to zero under steady-state operating conditions, even if the network output is subject to some error. Clearly, fuel ¯ow per unit time is dependent upon speed, but fuel mass per injection for a given cylinder air charge may be independent of speed. Thus, speed has not been used as an input to the neural network at present. Future work may, however, show that it is required to produce the required AFR accuracy. 4.2. Experimental methodology For engine test purposes the TFC neural networks were trained according to a procedure that avoids prior knowledge of intake port fuel transport characteristics. This method is described in detail in (Shayler et al., 1996), but may be summarised as follows. Only a basic understanding of the physical behaviour of the system is required, but as a result several approximations must be made. These necessitate the adoption of an iterative approach to determine the ®nal, correct TFC requirements. It was assumed that, since the TFC fuel is supplied to compensate for changes in the fuel ®lm mass, for a correctly compensated transient, the total change in fuel ®lm mass will be equal to the TFC fuel supplied. Also, changes to the mass of fuel entering or leaving the fuel ®lm at any given point in the transient, as a result of changing the injected fuel, are ignored. The latter of these two will give an inaccuracy in the calculated TFC requirement, but this will be less than that present in the TFC applied in the preceding test, and an iterative procedure should converge to the actual TFC requirement. By making these assumptions, the correction to the TFC requirement simpli®es to the dierence between the mass of fuel required to
155
give a stoichiometric mixture and the actual mass of fuel entering the combustion cylinder during the previous test. This actual mass was determined from the known cylinder air charge and the corresponding exhaust AFR. From this new TFC requirement, the corresponding fuel ®lm mass variation required as a network input was determined by integration with respect to injection number. The initial fuel ®lm mass was calculated from the relationship between the overall change in fuel ®lm mass and the change in air ¯ow rate for the training transients. Since three transient pairs were used, a quadratic relationship was required, with the assumption that the fuel ®lm mass was zero for zero air ¯ow. A simple plenum ®lling/emptying model of the inlet manifold, together with electronic throttle control, was used to predict air ¯ow at the time of fuel induction. Such a scheme is necessary to allow time for the fuel requirement to be calculated, injected and transported to the cylinder, for induction in phase with air ¯ow. 4.3. Summary of performance To facilitate the development and evaluation of the new fuelling control strategy, control of the fuel injectors was taken over by software running on an IBMcompatible 486 personal computer (PC). An exhaust AFR sensor was also installed to assess the transient fuel compensation accuracy. Electronic throttle control was provided by a positioning stepper motor driven by the fuelling software. The performance of the neuralnetwork-based TFC scheme was assessed for a range of transients. The experimental work was carried out on a 1.8 l Ford Zetec engine. The results of these tests are illustrated in Fig. 9, in terms of peak AFR excursion. The results include data for engine speeds of 1000, 1500 and 2000 rpm, for a network trained at only 1000 rpm. At each speed, the tests covered a range of air ¯ow changes both within and extending beyond the limits used to generate training data. From this ®gure it can be seen that the peak excursion is below 1.2 units in the majority of cases, there being only two transients for which the excursion is larger, but still less than 1.8 units, and these only occur when the lower normalised air ¯ow rate of the transient is below 0.25. These results compare well with those for the present production control system for the same engine, which gives typical excursions of up to 2.4 units with some as large as 4 units. The network-based system has a simpler structure and eliminates a number of look-up tables found in a more conventional system. There is little calculation required in addition to the propagation of data through the trained network. The relevant inputs (normalised air ¯ow, coolant temperature, and estimated
156
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157
found to be more successful. Addition and proportional correction schemes similar to those described for the previous application were examined (Goodman, 1996), but were found to be less successful, due to the increased complexity of the relationship between network input and output values.
5. Discussion and conclusions The introduction of microprocessor-based, electronic engine control units has enabled improvements to be made in many areas of systems control and monitoring. New applications and further improvements continue to emerge. The development of control strategy and, more generally, engine management software is an expanding task. This may contain literally hundreds of parameters which must be calibrated for each application. The expertise required is considerable, as are the resources vested in documenting and supporting applications. Set against this background, neural networks have attracted interest for reasons that include the possible simpli®cation of software development and calibration, and the utility of their pattern-classi®cation characteristics. As various investigators have reported, these have been exploited in a number of engine and vehicle systems applications. In the work reported here, three engine-related applications have been examined. These illustrate the diversity of problems and ways in which neural networks can be exploited. The results obtained using relatively simple networks and network training procedures have been generally encouraging. As important as the expertise in the neural-network ®eld, has been an understanding of the nature of the problem being addressed. The likelihood of a successful implementation depends on both.
Acknowledgements
Fig. 9. Summary of performance of neural-network-based compensation over a range of air ¯ow changes and engine speeds. The network was only trained on data from 1000 rpm tests.
fuel ®lm mass) require linear scaling from real units to valid network units, the converse being true for the TFC fuel mass output, the only other necessary calculation being to modify the estimate of fuel ®lm mass for the cylinder concerned by the addition of the calculated TFC fuel mass. For this application an absolute value network was
The authors wish to acknowledge and thank the Ford Motor Company for ®nancial support and for permission to publish this work.
References Aquino, C.F., 1981. Transient A/F control characteristics of the 5 l central fuel injection engine. SAE Technical Paper No. 810494. Bacon, A., Shayler, P.J., Ma, T., 1992. Potential for engine control using neural networks. IMechE Paper No. C448/057. Beaumont, A.J., Frith, A.M., 1994. Adaptive control of transient air±fuel ratio using neural networks. In: Proceedings of the 27th International Symposium on Automotive Technology and Automation (ISATA), Aachen, Germany. Davis Jr, L.I., Puskorius, G.V., Yuan, F., Feldkamp, L.A., 1992.
P.J. Shayler et al. / Engineering Applications of Arti®cial Intelligence 13 (2000) 147±157 Neural network modelling and control of an anti-lock brake system. In: Proceedings of Intelligent Vehicles '92 Symposium, Detroit. Feldkamp, L.A., Puskorius, G.V., Davis, L.I., Yuan, F., 1992. Neural control systems trained by dynamic gradient methods for automotive applications. International Joint Conference on Neural Networks 2, 798±804. Goodman, M.S., 1996. Transient air/fuel ratio control of spark ignition engines using neural networks. Ph.D. thesis, University of Nottingham. Hires, S.D., Overington, M.T., 1981. Transient mixture strength excursions Ð an investigation of their causes and the development of a constant mixture strength fuelling strategy. SAE Technical Paper No. 810495. Ina, T., Obayahsi, H., Shigematsu, T., 1986. Lean limit AFR control using engine speed variation. SAE Technical Paper No. 860413. Lenz, U., Schroeder, D., 1996. Arti®cial intelligence for combustion engine control. SAE Technical Paper No. 960328. Marko, K.A., 1991. Neural network application to diagnostics and control of vehicle control systems. In: Neural Information Processing Symposium (NIPS-91). Morgan Kaufman, New York. Mihele, W.P., Citron, S.J., 1984. An on-line roughness measurements technique. SAE Technical Paper No. 840316. Puskorius, G.V., Feldkamp, L.A., 1991. Decoupled extended
157
Kalman ®lter training of feedforward layered networks. In: International Joint Conference on Neural Networks, vol. I, IEEE, New York, 771±777. Puskorius, G.V., Feldkamp, L.A., 1992. Model reference adaptive control with recurrent networks trained by the dynamic DEKF algorithm. In: Proceedings of the 1992 International Joint Conference on Neural Networks, Baltimore. Puskorius, G.V., Feldkamp, L.A., 1993. Automotive engine idle speed control with recurrent neural networks. In: Proceedings of the 1993 American Control Conference, San Francisco. Rizzoni, F., 1987. Test transforms for rapid isolation of mis®ring cylinder. SAE Technical Paper No. 871915. Shayler, P.J., White, D.M., Wilkinson, A.C., Wakeling, R. 1990. The assessment of information sources for fault detection on spark ignition engines. IMechE Paper No. C422/006. Shayler, P.J., Goodman, M.S., Ma, T., 1996. Transient air/fuel ratio control of an S.I. engine using neural networks. SAE Technical Paper No. 960326. Shayler, P.J., Teo, Y.C., Scarisbrick, A., 1995. Fuel transport characteristics of spark ignition engines for transient fuel compensation. SAE Technical Paper No. 950067. Specht, D.F., 1991. A general regression neural network. IEEE Transactions on Neural Networks 2 (6), 568±576.