Determination by Neural Networks of the Optimal Spark Advance of a Highly Diluted Spark Ignition Engine M. Costa°, G.Palmesano*, G. Torella** °Istituto Motori, CNR, Viale Marconi, 8 – 80125, Naples – Italy (m.costa@im.cnr.it) *Second University of Naples, DIAM, Via Roma, Aversa (Naples) – Italy (g.palmesano@tin.it). **Italian Air Force Academy, 80078, Pozzuoli (Naples) – Italy. Abstract A neural network is designed to determine the optimal spark advance of a highly diluted spark ignition engine. The lack of a sufficient number of experimental data to be used as patterns for the phase of net training is overwhelmed by resorting to a simple physical model for the in-cylinder combustion phase. Although not able to simulate the complete engine behaviour, the model furnishes a number of virtual patterns to be used in addition to the real ones. The neural network is generated as backpropagating by optimising the number of neurons in the hidden layer and the transfer function. Accuracy of the network as a predictive tool for the spark advance maximising the brake torque is shown. The employed physical model for spark ignition combustion offers the way to extend the use of neural networks to a field where the high number of required patterns would require an effort of test bench activity unacceptable for a rapid vehicle prototyping. Keywords SI engine; optimal spark advance; neural network; learning; combustion model. Introduction Automated testing of engine control systems represents nowadays a critical factor for the competitive development of new vehicles. Present tendency is to substitute the calibration of the Electronic Control Unit (ECU) by maps with Hardware In the Loop (HIL) methodologies including algorithms for the simulation of the engine behaviour. This approach provides the proper coupling between input and output signals and engine parameters, although it is evident that the time needed for the simulation of the engine becomes a fundamental constraint for its employment. The present paper is devoted to explore the suitability of Neural Networks (NNs) to be used as engine simulators for ECU programming. At present, the high number of patterns needed for the phase of net training, makes for NNs to be of low interest for on-line applications as well as for preliminary ECU calibration. The need to perform experimental tests under different operating conditions would be excessively expensive in term of cost and time with respect to what needed by other techniques. Present works moves from the idea of generating virtual patterns to be used in addition to the real ones, recorded by operating a given engine on a test bench. The virtual patterns are constructed by employing the Vibe function, already used by authors to develop a model-based technique for ECU calibration [1]. The aim of the network is here to determine, as a function of the engine operating conditions, the optimal spark advance, namely the spark advance maximising the engine brake torque. The considered engine, 4 valves, 4 cylinders, 1.4 litres, is particularly......