Artificial neural networks for speeding-up the experimental calibration of propulsion systems

Artificial neural networks have proven to be very useful, in different areas of investigation, for the realization of black box models able to represent the relationship between input and output data with high precision. In the automotive sector too, they have found widespread use, particularly for sub-models involving non-linear aspects. Most of these applications are not predictive, but mainly exploit the ability to obtain useful input-output relationships trough learning on scattered data. In contrast, the novelty of this paper is the use of a neural network, trained with experimental data in a portion of the operation area of the system to be calibrated, within a method used for forecasting purposes. The proposed framework consists of an iterative cycle of experimental tests, with control parameters identified by the neural network model, to reduce the time required for calibrating the entire working plan. This type of approach requires knowing the areas where the predictions of the model are sufficiently accurate. Therefore, a study was conducted for three different types of internal combustion engines, to evaluate a general trend of the prediction accuracy of the model, as a function of the distance from the training zone. In particular, the area where the learning was performed (10% of the engine operating plan) was expanded by 200%, with an error of approximately 5 times that of the training phase. Finally, the proposed method was compared with the best fitting surface approach and was found to be significantly reliable and accurate.