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