Design of a Hybrid Time-Series Model for Fault Diagnosis and Degradation Prediction of Power Semiconductor MOSFETs

This paper presents a hybrid deep learning framework for fault diagnosis and degradation prediction of power semiconductor MOSFETs. The proposed model combines a multilayer perceptron for extracting static multi-parameter features with a long short-term memory network for learning temporal degradation patterns, forming a multi-task architecture that facilitates the simultaneous classification of device health states and prediction of future degradation trends. The input dataset integrates MATLAB/Simulink-based simulation data with the NASA Prognostics Center of Excellence MOSFET degradation benchmark dataset. Key degradation-sensitive parameters, such as threshold voltage, drain-source voltage, drain current, on-state resistance, body-diode voltage, and junction temperature, are used as input features. A health-score metric is utilized to classify device states into normal, warning, and fault categories. Validation results demonstrate classification accuracies of 98% and 95% on the validation and test sets, respectively. Overall, these results confirm that the proposed framework effectively captures degradation characteristics and provides a viable foundation for predictive maintenance of power semiconductor devices.