Estimating Ignition Timing from Engine Cylinder Pressure with Neural Networks

A study was conducted to determine the ability of neural networks to extract high level control information from cylinder pressure data. Various experiments were performed us ing neural networks for pattern recognition on a series of data files consist ing of cylinder pressure versus crank angle. The goal of these experiments was to estimate spark timing based on the cylinder pressure signature -all other engine parameters were held constant during the data collection process. Test results indicate that an approximate spark timing value can be obtained using cylinder pressure data as the inputs to a neural network and spark timing as the output. 1 . I n t r o d u c t i o n In-cylinder pressure data provides one of the most direct measures of combustion quality in an internal combustion engine. Cylinder pressure data has been used for design and diagnostic purposes since the IC engine was developed. Cylinder pressure and volume data can be used to calculate engine torque, indicated mean effective pressure (IMEP), indicated efficiency, bulk temperature, burn rate and heat release. Statistical analysis of the same data can provide informat ion about combust ion variability. Recently, more attention has been given to the use of cylinder pressure for real-time engine control [Pes89, KSS88, AP87, GP89, HA86]. Pressure-based engine control techniques require that information be extracted from cylinder pressure data. Examples of pressure based engine control include: • The location of peak pressure can be monitored to set the optimal spark timing [AP87, Pes89] for power or fuel economy . • A/F ratio can be estimated from pressure data [GP89]. This could conceivably be used for feedback fuel con t ro l . • Indicated mean effective pressure (IMEP) can be calculated from the pressure data and maximized using a h i l l c l imb ing o r g r ad i en t de scen t technique [DL51, Kal58, Bla62, Flo64, KPF89, SW90] to achieve maximum t o r q u e . Pressure-based engine control has not been practical for two reasons. The piezoelectric pressure sensors cur ren t ly used a re expensive, fragile, temperature sensit ive and requ i re a c lose -coup led charge amplifier to produce a high level signal. From a s ignal process ing s tandpoint , calculation of anything more than peak pressure has been too slow for use in real time control. Recent developments in sensor technology offer some hope on the hardware side of the problem. A compact piezoelectric sensor with a charge amplifier and temperature compensation built in has been developed [AP87, Pes89] although, there are no immediate plans to commercial ize the sensor. Piezoelectric washers which are installed between the spark plug and cylinder head have been shown to provide qualitative pressure data at relatively low cost and are reportedly in use in at least one production vehicle for knock sensing. Within the last few years, there have been advancements in the development of low cost fiber optic pressure sensors which are suitable for engine work, and could be mass-produced at low cost. In many ways, hardware developments appear to be outpacing developments in the area of s ignal processing. Notable exceptions include the adaptation of digital s ignal processors (DSP) for realt ime calculation of IMEP from pressure data. An investigation was therefore undertaken to assess new signal processing techniques for use in pressure-based engine control. 2 . The Role of Neural Networks It is our premise that advanced signal processing techniques being developed in other fields may be adaptable for pressurebased engine control. As a first step in this investigation, a study was conducted to determine the potential of neural networks to discern informat ion f rom cyl inder pressure data. Rapid advancements in the development of neural network ICs suggest the possibility that they could be costeffective in selected engine applications in the near future. The specific focus of this study was to demonstrate the ability of neural networks to extract ignition timing from a cylinder pressure wave form. It is understood that there are easier ways to determine ignition timing. However, this study represents a first step in the development of neural networks which can recognize ignit ion timing, fuel/air ratio, and other parameters. The eventual goal of the project is to correlate engine data (cylinder pressure d a t a , m a n i f o l d p r e s s u r e , e n g i n e temperature, engine speed) with engine emissions, and use this information in a neural network based engine controller. 3 . Cylinder Pressure Data Cylinder pressure data for the pattern recognition experiment was obtained from SuperFlow Corporation. The pressure traces were generated on a 350 in.3 Chevrolet engine at 4000 RPM. The data was captured with SuperFlow's Engine Cycle Analyzer (ECA). The original files contained cylinder pressure (measured in "bar") at crank angles from 180o before top dead center (BTDC) to 180o after top dead center (ATDC). Each file contained 720 data points at 0.5 degree crank angle spacing. Data was taken at various different spark timings from 5o to 35o BTDC in 5o increments. In addition, data was taken at a spark timing of 22o BTDC to provide a test case for the pattern recognition experiments. The engine speed, air/fuel ratio, and throttle setting were all held constant throughout the data collection period. Thirty independent sets of data for each spark timing were taken over a period of approximately 45 seconds. The total elapsed time to collect all the data, changing only the spark timing between sets of data, was 27 minutes.