Potential of Machine Learning Methods for Robust Performance and Efficient Engine Control Development

摘要

Increasingly strict legislation for greenhouse gas and real-world pollutant emissions makes it necessary to develop fuel-efficient and robust control solutions for future automotive engines. Today’s engine control development relies on traditional map-based and model-based control approaches. Due to growing system complexity and real-world requirements, these expert-intensive and time-consuming approaches are facing a turning point, which will lead to unacceptable development time and costs in the near future. Artificial Intelligence (AI) is a disruptive technology, which has interesting features that can tackle these challenges. AI-based methods have received growing interest due to the increasing availability of data and the success of AI applications for complex problems. This paper presents an overview of the state-of-the-art in Machine Learning (ML)-based methods that are applied for engine control development with focus on the time-consuming calibration process. The overview here shows that the vast majority of studies concentrates on regression modelling to model complex processes, to reduce the number of model parameters and to develop real-time, ECU implementable models. The identified promising directions for future ML-based engine control research include the application of reinforcement learning methods to on-line optimize engine performance and guarantee robust performance and unsupervised learning methods for data quality monitoring.