Neuronal networks for process optimization and the diagnosis of internal combustion engines

Preliminary computations of process results are being made on the "virtual engine" from the very start of the development process of combustion engines. These computations are based on simple physical models. The heat release function is modelled by a neural network, which takes into account numerous influencing parameters, such as for diesel engines with common rail injection in passenger vehicles. The data used to train the network is taken from indicated pressure measurements, by means of which heat release functions are established. Apart from this neural networks are applied to present functionl processes in the cylinder, when drive systems are being simulated with combustion engines and more complex supercharging and control systems. Thus the computation rate is increased and the chance of an on-line process optimization with the engine running is opened. Results of simulation computations are being demonstrated. The pressure curve is best suited to evaluate the process within the cylinder, since it shows the highest information density while using the simple models mentioned above. Since the pressure curve is normally not available on series engines, it is generated from the signal of a vibration sensor as it is used for knock detection. The transfer function between vibration and pressure signal is demonstrated in a neural network for the overall operating range.