Remaining Useful Life Prediction of Power MOSFETs Based on Deep Reinforcement Learning

The power metal–oxide–semiconductor field-effect transistors (MOSFETs) are the critical components extensively utilized in various fields. To ensure the safety and reliability of MOSFET-based systems, the remaining useful life (RUL) prediction techniques can be applied to calculate the RUL during the early to middle stages of equipment degradation, thereby avoiding more severe failures. Over the past few years, the RUL prediction technology utilizing reinforcement learning (RL) has steadily matured, with the deep deterministic policy gradient (DDPG) algorithm demonstrating excellent predictive performance in time-series data prediction. However, the complicated degradation mechanisms of power MOSFETs and the limited historical degradation data make it challenging for the hyperparameters of DDPG model to effectively converge to the global optimum. Consequently, the DDPG-based prediction method suffers from poor generalization and adaptability, leading to larger prediction errors. To address the issue, we propose a novel prediction method, named GDDPG, which integrates Gaussian process regression (GPR) with DDPG. First, monitoring data of MOSFETs are used to construct state matrix and optimal action space. Second, during the iterative prediction phase, the GPR algorithm is introduced to correct each single-step prediction result from the DDPG model. Finally, the corrected prediction value is incorporated into the next iteration’s state variables for subsequent prediction. Validation on the publicly available NASA prognostics center of excellence (PCoE) dataset “MOSFET Thermal Overstress Aging” demonstrates that the proposed method consistently surpasses all comparison methods across multiple evaluation metrics. Thus, the GDDPG-based method significantly enhances the precision and dependability of RUL prediction for MOSFETs.