Hybrid Deep Learning for Si MOSFET Fault Diagnosis and Prediction

Si MOSFETs are widely used in power conversion systems; however, long-term operation under repetitive switching and electro-thermal stress leads to progressive degradation and eventual failure. Two representative failure modes are commonly observed: gate-oxide degradation and packaging-related degradation, which often exhibit different evolution patterns. This paper proposes an AI-based diagnosis and prognostics framework that jointly leverages steady-state time-series information and fixed-length features extracted from turn-off transients. The study utilizes the NASA Open Accelerated-Aging dataset and reorganized/preprocessed data supported by MATLAB/Simulink measurement cir-cuit modeling. Physics-informed rule-based labeling is applied to discriminate normal, gate-oxide, and packaging-related conditions based on degradation indicators such as Rds_on evolution. The trained model is further interpreted via permutation importance to quantify whether gradual/abrupt degradation indicators and transient features contribute to decision-making. Performance is assessed on held-out tests and synthesized cases sampled from baseline operating distributions to examine consistency under previously unseen conditions.