Automotive Feature Coordination based on a Machine-Learning Approach

The number of cyber-physical features within modern automotive powertrains is continuously rising. Such features are often not independent from each other, but interact. Such interaction is either known & desired or unknown & undesired. We propose a centralized coordination of such automotive features based on a machine-learning approach. After the generation of training data, learning of a neural network takes place during design-time. Embedded source code of the learned and frozen net is then automatically generated for the electronic control unit. A machine-learning-based feature coordinator for a gasoline combustion engine was prototypically developed. We applied hyperparameter optimization to find a minimal sized model with high accuracy. The trained neural network was successfully transferred to a combustion engine control unit.