Cost Efficient Training Method for Artificial Neural Networks based on Engine Measurements

Abstract

The artificial intelligence is an accurate predictive tool for different kinds of internal combustion engine (ICE) applications. However, the training process can be expensive due to the high computational and measurement costs. This work aims to describe a general methodology that can be applied to cost-efficiently train multilayer perceptron type artificial neural networks with measurement data from ICEs. The created methodology is based on analyses of a high-resolution dataset measured on a commercial diesel engine. Different methods and recommendations are presented for the model creation, evaluation, training method selection, input feature selection and architecture selection. In addition, a method is described in order to select the appropriate measurement resolution that provides proper information for training with minimal fuel consumption. The investigation showed that the presented workflow can reduce calculation time and fuel consumption, while maintaining good model accuracy. The method can be applied for any ICE related artificial neural network problems, but it can also be an aide for other research fields.