Machine Learning-based Condition Monitoring from Power Cycling Data for Silicon Carbide Power Inverter Modules

With the rise of electric vehicles, condition monitoring of power transistors is essential to prevent failures. This study investigates machine learning-based models for estimating the remaining useful life (RUL) of silicon carbide <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(\text{SiC})$</tex> power transistors by forecasting the normalized forward voltage drop increase until end-of-life. Utilizing extensive power cycling data from <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$\text{SiC}$</tex> inverter module qualification, we evaluate four state-of-the-art long-term series forecasting neural network architectures: LSTM, Patchmixer, PatchTST, and DLinear. The analysis reveals that the PatchMixer and PatchTST architecture demonstrate superior performance, achieving RUL estimation errors of around and below 5%. The study demonstrates that optimizing the input configurations improves the accuracy of the models while reducing their size. Application of the models to degradation patterns outside of the training distribution has revealed a lack of generalization capability, indicating the need for data that resembles practical application conditions. Future work should focus on evaluating model implementations for embedded hardware.