Systematic Comparison of Supervised Learning Methods to Reduce Calibration Effort in Engine Control Development

Complexity of engine control systems is continuously growing due to an increased number of subsystems and the need for robust performance. For traditional map-based as well as state-of-the-art model-based approaches, this will lead to unacceptable development costs and time, for future engines. Parametrization of the embedded models using supervised learning regression methods can immensely reduce the number of calibration parameters and hence the calibration effort. However, a methodology for performance assessment of different promising data-driven modelling methods for engine control development is currently lacking. In this paper, a systematic methodology that assesses model inaccuracy, and also implementation aspects such as calibration effort and computational complexity is introduced. This method is applied to assess the potential of SL methods for parametrizing the feedforward controller of a modern Diesel engine air-path controller. Using requirement analysis and the specified performance criteria, two regression methods were selected: artificial neural networks (ANN) and support vector machines (SVM). After careful data selection and model training, performance is compared with the benchmark controller, which uses a physics-based model. From simulation results, it is shown that a 97% reduction in the number of calibration parameters with both regression models can be realized. For a standard test cycle, cumulative engine out NOx emissions with regression based controllers are close to the allowable inaccuracy of 10% compared to the benchmark controller. Among the two methods, ANN shows the best performance for the studied performance criteria of inaccuracy, number of calibration parameters and computational complexity.