Jun 10, 1998 - emissions model within an optimising routine running as part of the vehicle ... A simulation was required to investigate the effect of control strategy changes on the engine ..... Figure 6 - Search method for locating IOL. 0. 20. 40.
Prediction of Diesel Engine Exhaust Emissions using Artificial Neural Networks Chris Brace, University of Bath
Neural Networks in Systems Design IMechE Seminar S591 Held at: Lucas Electrical and Electronic Systems, Solihull on the: 10 June 1998
Page 1 of 11
Summary Three applications of neural networks to the prediction of Diesel engine fuel consumption and emissions have been developed. The uses are quite distinct although they all used the same experimental data as their basis. One network was successfully trained to predict transient changes in emissions levels following rapid changes in engine operating condition. The second was used to predict emissions during a legislative drive cycle. The final example presented was used in a powertrain controller to identify the ideal set point for engine speed and load to minimise fuel consumption and emissions during steady driving. The design and training of each network is discussed and results presented from each application. Some general comments on suitability and ease of use of the technique are included.
1. Introduction Although neural networks have been used with varying degrees of success for engine modelling (1), control (2 & 3) and fault diagnosis (4) their use in production vehicle applications is still quite rare. Nevertheless, the characteristics exhibited by neural networks can be useful in this field. A series of neural networks have been developed to predict various aspects of Diesel engine emissions production in support of an integrated powertrain control project (5). The powertrain under investigation comprised a modern high speed direct injection Diesel engine and a continuously variable transmission (CVT). The three main applications of neural networks were: • • •
combustion model within a transient engine model to aid detailed control strategy development emissions model for drive cycle emissions prediction to aid powertrain control strategy development (6) emissions model within an optimising routine running as part of the vehicle powertrain controller (7)
2. Transient Emissions Prediction Network 2.1. Simulation Design A simulation was required to investigate the effect of control strategy changes on the engine behaviour, particularly emissions performance. A typical manoeuvre to be controlled was the 'step up' from low to high pedal demand at constant engine speed. This was to be investigated over relatively short periods of around ten seconds of real time. An engine simulation was constructed within a modular environment - Bathfp (8). Each subsystem of the engine was represented by a discrete model, connected to the surrounding models within a graphical user interface as shown in Figure 1. It was important that the model ran quickly as strategy changes needed to be investigated quickly during the development process. Most of the sub-models used analytical or simple empirical algorithms. In the case of the engine combustion model the requirement was to predict engine torque, air mass flow, exhaust temperature and the emissions - smoke, hydrocarbons (HC), oxides of nitrogen (NOx) and particulates (PM). The phenomena involved would have been very difficult to represent analytically with sufficient accuracy and speed of calculation, requiring an alternative technique to be sought. The approach adopted was that if the input conditions to the combustion chamber are known then the output emissions will be predictable. The validity of the approach is dependent on the input conditions being fully described. This necessitates a multi input empirical model. The neural network is a useful method of constructing a multi-input, multi-output empirical model which runs quickly. The inherent smoothing capabilities of the network were considered advantageous in comparison with simple interpolation methods where potentially noisy experimental data are used. It was also considered of little relevance whether the combustion event in question is one of a series of similar events, as seen in a steady state test, or one event forming part of a transient. This Page 2 of 11
approach allowed the use of steady state experimental data in a transient model. Steady state data are much cheaper to generate and are likely to be more repeatable. The essential pre-requisite is sufficient data to represent the system accurately. If suitable hardware was not available the data may be generated using a more traditional analytical model running off-line although calibration of this model may prove problematic. In this case a suitable engine was available on a steady state test rig. 2.2. Data Generation A typical engine mapping exercise involves running at a series of speed/load points to fully cover the expected range of operation. In order to represent the behaviour of the engine during transient manoeuvres the engine was additionally tested at conditions not normally encountered during steady state running. In a turbocharged Diesel engine the effect of reduced boost pressure is particularly important for transient operation. The test work is summarised in Table 1 below. Data Set
Independent Variable under Investigation
Value of Independent Variable
None - design point map None - design point map - reduced no of points Injection timing Injection timing Charge temp low Charge temp high Boost pressure low Boost pressure high Coolant temp very low Coolant temp low Coolant temp high EGR fraction zero EGR fraction low EGR fraction high
all on nominal all on nominal
No of Speed x Load Points 8x13 4x4
3 degrees earlier 3 degrees later o 30-50 C o 70-107 C approx. 0kPa 60-70kPa o 25-30 C 63C (83 is norm) o 93-100 C 0 6% less than normal 12% more than normal
4x4 4x4 4x4 4x4 4x4 1x2 4x4 4x4 4x4 4x4 4x4 4x4
3 4 5 6 7 8 9 10 11 12 13 14
Table 1 - Test data points for training network 2.3. Network Design and Training The network outputs were determined by the predictions required for strategy development. These consisted of the emissions together with fuel consumption and torque output. The inputs were those engine variables seen as having a direct relevance to the outputs. It was important, as with any network design, to include sufficient inputs to fully describe the system but to avoid the inclusion of irrelevant inputs as this would introduce the likelihood of spurious relationships being identified. An understanding of the physical process is essential at this stage to separate causality from coincidence. Stochastic back propagation was used to train the network, the number and size of the hidden layers being determined empirically according to the error value and stability observed. A single hidden layer containing 30 neurons proved the best solution for the application. 2.4. Results In order to test the network response alone a series of transient tests were performed on a dynamic test bed using the same engine as for data generation. The relevant input variables were monitored and passed to the network, resulting in predictions of emissions and performance. Figure 3 shows a comparison of some of these predictions with the measured experimental data. The test under consideraton is a rapid change in load and speed from an initial 15Nm at 2000rev/min to 120Nm at 3500Nm after 5 seconds and then back to the initial condition after 30 seconds.
Page 3 of 11
The predictions in the initial state are reasonably accurate, the NOx prediction being closest. Immediately following the step change the network predicts a sharp rise in NOx production. The experimental data were generated using conventional emissions analysers and as such exhibit the slow response times expected. The pure transport delay has been removed, leaving the effect of the slow rise time of the instruments. Any rapid spikes will be attenuated in the instrumentation, leaving the network prediction as the more realistic when the rapid change in load is considered. A similarly rapid fall at the step down is also attenuated by the instrumentation. There is an appreciable error in the predicted NOx level at the higher power condition, which is reduced slightly over time as the measured level increases slowly. This is due to thermal effects not considered in the model such as warming of the combustion chamber at the high power condition. Hence the network predicts relatively constant NOx production over the 25 seconds whereas the real system increases its NOx output with temperature. If more representative data, such as valve bridge temperature, were available this effect could be repeated in the network. The network had only inlet air and coolant temperature available as inputs, both of which are only weakly related to combustion chamber temperature. The HC traces show much of the same characteristics of the NOx, although the instrument is known from previous work to have a slower response than the NOx analyser used. This is evident from the experimental trace. The HC levels at the high power condition fall over time due to the slow thermal response of parts of the engine having the opposite effect on the HC production mechanisms. Smoke is measured experimentally using a full flow optical opacimeter with a very fast response and no transport delay other than the transit time of the gas along the exhaust system. The experimental trace thus matches the network prediction well with a slightly broader peak following the step up due to mixing within the exhaust system. The discrepancy in absolute values arises from the difficulty in converting opacity measurements to the Bosch smoke units measured on the steady state cell with which the network was trained. The steady state device uses a reflective paper method unsuited to transient testing. In summary, the network predictions following rapid changes in operating conditions matches the expected response of the engine. The instrumentation used here to measure NOx and HC introduces a delay to the experimental data making accurate comparison difficult. The optical opacimeter used to measure smoke transiently makes comparison easier and shows the validity of the approach. The network was used successfully to predict the effect of control strategy changes on transient emissions production, particularly smoke.
3. Drive Cycle Emissions Prediction Network A common test used in the evaluation of competing control strategies is the European drive cycle (ECE15 + EUDC) which consists of 1200 seconds of simple acceleration, cruise and deceleration sequences. A simulation was required to evaluate competing strategies over this cycle at the concept stage. Evaluation criteria were emissions (HC, NOx and PM) and fuel consumption. The simulation needed to be very fast since a large number of strategies needed to be compared within a short period. Since the cycle is predominantly steady state with only modest transients a more simple engine model structure was proposed than the transient model described above. In addition to engine speed and load, the only other engine model input was coolant temperature. It was judged that the remaining input variables would reach their design points rapidly and have no noticeable effect when compared with steady state running.
Page 4 of 11
3.1. Simulation Design The approach adopted was to construct a simple instantaneous model of the controller and powertrain within an Excel spreadsheet with predictions made at one second intervals throughout the test. This gave an adequate representation of the vehicle dynamics and lent itself well to reporting and integration with other applications. The simulation predicted engine speed and load and combined them with typical experimental coolant temperature to provide input data to the engine model. A neural network was used for the engine model as experience with the transient model had suggested that it was a useful tool for the task. 3.2. Network Design and Training The number of network inputs and outputs were again determined by the application. A network of the form shown in Figure 4 was trained using a subset of the data gathered for the transient model above. It was important to include only data from tests where the network input variables had been the independent variable. Including other data would have broken the relationship between network inputs and outputs and adversely affected network accuracy. The data used are summarised in Table 2 below. Data Set 1 2 9 10 11
Independent Variable None - design point map None - design point map - reduced no of points Coolant temp - very low Coolant temp - low Coolant temp - high
Value of Independent Variable all on nominal all on nominal o
25-30 C o 63 C o 93-100 C
No of Speed x Load Points 8 x 13 4x4 4x4 4x4 4x4
Table 2 - Test data points for training network The best accuracy of training was achieved with 30 neurons on a single hidden layer. Using two hidden layers introduced a tendency to instability during the training process with no increase in peak accuracy. The errors when compared to the training data for the network used are shown below. mass HC g/hr 2.78
mass NOx g/hr 4.43
mass PM g/hr 5.42
Bosch Smoke 6.26
fuel flow kg/hr 2.53
Table 3 - Network Training Errors - RMS Errors on training data (% full scale) 3.3. Results The simple simulation with neural network engine model was used successfully to investigate a wide range of control strategies and hardware configurations. Of particular interest was the ability to evaluate a candidate solution against the 'ideal' case. This ideal case was a simulation where no constraints over engine operating point were imposed. This would correspond to a perfectly efficient continuously variable transmission with no ratio constraints. Hardware limitations were then introduced progressively to identify areas where improvement could have the greatest benefit. Another use was to predict the effect of different engine operating lines (see section 4) on emissions production. Figure 5 shows a portion of the European drive cycle (the extra urban drive cycle or EUDC) showing predictions for NOx production and fuel consumption from two different controller calibrations. The plot shows clearly the trade off between NOx and fuel consumption. A controller which is well optimised for one will be poor at the other and vice versa. This correlates very well with experimental findings (7).
Page 5 of 11
4. Operating Point Optimiser During the course of the integrated powertrain control project a number of strategies were developed which required knowledge of the ideal operating line (IOL) for the engine. This was defined as the operating line which allowed the engine to produce any desired power whilst returning the lowest possible fuel consumption and emissions. There is no single line that can minimise all of the emissions and fuel consumption simultaneously as each pollutant is formed by a different mechanism. There are various solutions to this difficulty, the simplest being to optimise for one pollutant at a time. This allows regions of operation where NOx, for example, is a known problem to be treated differently from those where HC emissions are poor. Another approach is to combine the pollutants in a weighted sum which is then minimised. This allows the ranking of pollutants in relation to legislative limits otr other criteria. In order to demonstrate this approach a C function was written to determine the IOL on-line within the powertrain controller. 4.1. Design Figure 6 shows graphically the steps taken to determine the IOL. Initially a straight ‘optimum’ line is drawn between the fixed points of minimum power (at idle speed) and maximum power (at maximum engine speed). The line is defined by a series of points spaced at 5kW intervals from minimum to maximum engine power. The engine model is used to predict the exhaust emissions at point 1 where the initial line crosses the 5kW constant power curve. The predictions are normalised with individual scale and offset factors over the range 0 to 1 to allow direct comparison between values of widely differing magnitudes. The various exhaust emissions are weighted according to the values set during calibration and summed. This step is repeated at points 2 and 3 which lie on the same power curve but at lower and higher engine speeds respectively. The point with the lowest weighted sum is chosen as the new ‘ideal’ point for 5kW. This process is repeated for each power step up to 95% full power. The procedure is repeated at the next controller iteration, which will move the line again in the direction of reduced exhaust emissions. If the ideal point moves in the same direction twice consecutively the speed step between points 1, 2 and 3 is increased. If the middle point is chosen twice in succession, or if the optimum point is approaching the outer envelope of engine performance, the speed step is reduced. The step size is also reduced if the algorithm appears to be hunting around the solution. This simple procedure quickly settles on an ideal line for the chosen weightings, helped by the simple shapes of the exhaust emissions maps. If the surfaces were more convoluted and incorporated local minima a more complex optimisation routine would be required (9). Once settled, the line will change slowly with fluctuations of engine water temperature and exhaust emissions weightings. 4.2. The Engine Model The model used to predict the exhaust emissions in the operating line optimiser is an empirical model developed using the same experimental data as used for the drive cycle simulation network. The only difference is that the emissions are presented in specific units of g/kWhr which minimises their variation across the operating range. These data were used to train a neural network of the same form as that shown in Figure 3. Here a single hidden layer containing 7 neurons gave the lowest training error and resulted in a compact network for fast interrogation when included in the controller. Table 4 below shows the errors in network predictions when compared to the training data. Figure 7 shows the network predictions for NOx at 85oC water as a wireframe surface with the experimental data superimposed. The fit can be seen to be good subjectively.
Page 6 of 11
spec. HC g/kWhr 7.80%
spec. NOx g/kWhr 6.74%
spec. PM g/kWhr 4.64%
Smoke Bosch 9.87%
BSFC g/kWhr 2.43%
Table 4 - Optimiser network training errors on training data (RMS of percent full scale) 4.3. Results Figure 8 shows a series of ideal operating lines produced using the function where one pollutant at a time has been optimised. This is useful as it allows simple inspection of the compromises inherent in powertrain calibration for economy and emissions. The fuel economy line lies very close to the limiting torque curve (LTC) of the engine. This results in very low engine speeds being selected for a given power demand. This has some advantages for noise and refinement as well as fuel consumption but is particularly bad for NOx production as discussed in section 3.3. It can be seen that the IOL for NOx is well below the LTC, leading to relatively high engine speeds for the power demanded. This has a beneficial effect on drivability as there will always be some useful power reserve available by moving up to the LTC without the need to change engine speed. The IOLs for HC and PM are very similar to one another and represent a good compromise between the two extremes of the IOL for fuel consumption and that for NOx. The optimiser function was successfully integrated into a powertrain controller which demonstrated ease of tuning via user defined weights. The emissions performance of the vehicle could be adjusted to deliver the emissions required while the adverse effects on fuel consumption could be predicted and assessed to allow the optimum compromise to be reached.
5. Lessons Shortcomings of the neural network based approach may be split into two areas, errors in the trained network and inconvenience in their training and implementation. 5.1. Errors The principal inaccuracies in the models arise from three sources. The first is the lack of appropriate input variables to the model. In particular the transient emissions model really needs a measure of combustion chamber temperature. Suitable instrumentation was not available on the test engine. Measurement of valve bridge temperature or a similar variable would have allowed better representation of the thermal history of the engine following a step change. To achieve the best results tests would need to be performed to generate training data with valve bridge temperature as the independent variable. This may be difficult to arrange without undue disturbance to the other input variables. The second source of inaccuracies is scarcity of data. Although almost three hundred separate test points were visited to generate training data some areas of transient operation were sparsely represented. This makes precise representation of trends difficult, even with the good interpolation properties of the neural network. Modern automated test cells offer the capability to run vast numbers of data points relatively simply although all the independent variables under consideration must be under automatic control to derive the full benefit. It should be stressed that both of the above sources of error are significant independently of the technique used to build the empirical model. The neural network may perform better than some alternatives in the case of sparse data. A third distinct source of error is due to the smoothing inherent in the network. If a discontinuous or highly convoluted function is being modelled the network will require large numbers of neurons and/or hidden layers and may never pass through peaks as accurately as an interpolation routine. Page 7 of 11
5.2. Inconvenience The manual, iterative technique used to set the number of hidden layers and neurons on each proved a little time consuming initially although with only modest experience the training process can be speeded up considerably. In all cases it is advisable to inspect the finished network predictions carefully to check that the learnt behaviour is realistic. Surprisingly good training errors can be returned by networks which in no way represent the desired relationship. In comparison with a linear interpolation approach the technique may initially seem to have significant setup time penalties. In fact, this is not always the case. If experimental data are used for the interpolation it is crucial that the raw data are carefully inspected to allow outliers to be removed by hand. This stage is less important when training a network as the network will tend to smooth out these effects.
6. Conclusions Three applications of neural networks to the prediction of Diesel engine emissions have been developed and used successfully. All three models were trained to recognise the patterns linking inputs and outputs of the engine using steady state experimental data. Network design and training was straightforward and rapid, allowing powerful empirical models to be constructed rapidly. Integration with the powertrain controller code and engine simulation was simple due to the use of the ANSI C programming language throughout. Advanced controller prototyping tools often include a neural network capability to further simplify their application. The network structure gave a useful smoothing effect to the model since the experimental data were subject to normal random errors. This had the effect of making the errors appear larger than may be expected if training using data generated using a simulation but the surfaces represented by the networks have a good subjective shape and fit to the data. This smoothing effect gives the network useful advantages over a simple linear interpolation model in this case but may be a problem where highly convoluted relationships are modelled.
Acknowledgements The work reported in this paper was undertaken as part of a research program supported by the Department of Trade and Industry, Ford Motor Co, Lucas Diesel Systems and Johnson Matthey Catalytic Systems. The author is grateful for the assistance and contributions of the collaborators. Particular acknowledgement is due to fellow researcher, Dr Mike Deacon, and support staff at the University.
References 1. Shayler P J, Darnton N J, Ma T, Predicting the Fuel Consumption of Vehicles for Drive Cycles Starting From Cold Ambient Conditions, EAEC 5th International Congress 1995 SIA9506A27 2. Bacon A, Shayler PJ, Ma T, Potential for Engine Control using Neural Networks. IMechE 1992 C448/057 3. Morita S, Optimization Control for Combustion Parameters of Petrol Engines Using Neural Networks - In the Case of On-Line Control, International Journal of Vehicle Design Vol 14 nos 5/6, 1993 4. Scaife MW, A Neural Network for Fault Recognition. SAE Congress Session 1 p23 Controls for Engines 1993 Page 8 of 11
5. Brace CJ, Deacon M, Vaughan ND, Burrows CR, Horrocks RW, Integrated Passenger Car Diesel CVT Powertrain Control for Economy and Low Emissions. IMechE International Seminar S540 'Advanced Vehicle Transmissions and Powertrain Management' 25 -26 Sept 1997 6. Deacon M, Brace CJ, Vaughan ND, Burrows CR, Horrocks RW, Impact of Alternative Controller Strategies on Emissions from an Integrated Diesel CVT Powertrain. Submitted to Proceedings of The Institution of Mechanical Engineers Journal of Automobile Engineering (Part D) 1997. 7. Deacon M, Brace CJ, Vaughan ND, Horrocks RW, Burrows CR, An Operating Point Optimiser for the Design and Calibration of an Integrated Diesel/CVT Powertrain. Submitted to Proceedings of The Institution of Mechanical Engineers Journal of Automobile Engineering (Part D) 1997. 8. Deacon M, Brace CJ, Guebeli M, Vaughan ND, Burrows CR, Dorey RE, A modular approach to the computer simulation of a passenger car powertrain incorporating a Diesel engine and continuously variable transmission. IEE International conference on Control (1994) University of Warwick. 9.
Donne MS, Tilley DG, Richards CW, The Use of Multi-Objective Parallel Genetic Algorithms to Aid Fluid Power System Design, Proc. IMechE Vol 209, I01894, 1995
Figure 1 - Bathfp circuit of transient engine model
Figure 2 - Schematic of transient engine model network
Page 9 of 11
Transient Rig with Predictions from Net7 - Step Changes cjb04c.opj
Simultaneous Speed and Fuel Step
800 Experimental Predicted
0 HC (ppm C3H8)
0 6 4 2 0 0
30 Time (sec)
Figure 3 - Transient Emissions Predictions compared with Experimental Data 250
Iol for NOx IOL for Fuel consumption
200 150 100 50 0
Fuel flow (kg/hr)
10 8 6 4 2 0 800
Figure 4 - Neural network used for emissions prediction in drive cycle simulation
Figure 5- NOx and fuel consumption predictions during European Drive Cycle
Page 10 of 11
Torque (% of Maximum)
40 LTC Final Ideal Line Initial 'Ideal' Line Contours of Emissions Lines of Constant Power
steps along constant power curve
Engine Speed (% of maximum) Figure 6 - Search method for locating IOL Optimum Lines Predicted by OP99.c Water 100C
14 12 NO x (g/kW hr)
OPT_DEM3.org Feb 97
8 Torque (Nm)
6 4 2
Limiting Torque Curve Ideal Line for NOx Ideal Line for Fuel Economy Ideal Line for HC Ideal Line for Particulates Lines of Constant Power
d( r ev
(N q ue
Engine Speed (Rev/min)
Figure 7- Neural Network representation of NOx data At 85oC water temperature, experimental data superimposed
Figure 8 - Ideal operating lines predicted using neural network engine model
Page 11 of 11