Timely Predicting Power Mosfet Failure in EV: AI Based Metodology and Deployment at the Edge

Predicting the residual useful life (RUL) of power MOSFETs in electric vehicles (EVs) is crucial due to their role in transferring power from batteries to the engine. Despite its practical importance, most solutions are theoretical. This paper proposes a methodology for MOSFET RUL prediction tailored for edge devices like Nvidia Jetson ORIN and STM32. Two machine learning-based models are introduced. The first identifies the Rds(on) degradation law from a dataset that best matches the observed degradation, suitable for low-noise scenarios but lacking fine-tuning. The second model, using an autoencoder with dual decoders, is more robust. One decoder ensures the Rds(on) degradation reflects normal device behavior, while the other provides timely warnings as Rds(on) approaches critical limits, enabling proactive replacement. The paper validates these methods with theoretical and experimental datasets, emphasizing the second model's successful implementation on edge devices, demonstrating its practicality for EV applications.