Machine Learning Based Soft Sensor Development for automotive fuel cell systems

The paper discusses the development of a machine learning based soft sensor for polymer electrolyte membrane fuel cell systems (PEMFC). Different modelling techniques such as decision trees, artificial neural networks (ANN), linear regression, and support vector machines (SVM) were implemented and compared. Artificial neural networks correctly predicted the cathode humidity condition of the fuel cell with high accuracy for Federal Test Procedure (FTP-75), Worldwide harmonized Light vehicles Test Procedure (WLTP) and New European Driving Cycle (NEDC) driving cycles, outperforming other machine learning models in terms of RMSE (root mean squared error) and R-squared (coefficient of determination) values, which are the metrics used for performance evaluation of regression models. Furthermore, standard automotive driving cycles, i.e., FTP-75, WLTP and NEDC, are evaluated for their suitability in generating the training data for machine learning applications. Machine learning models trained on data generated by simulating WLTP and NEDC driving cycle produced the best results in terms of consistency and accuracy.