Double-Phase Adaptive Neural Network for Condition-Based Monitoring of p-GaN HEMT Under Repetitive Short-Circuit Stresses

Recently, p-GaN gate high electron mobility transistors (HEMTs) have emerged as competitive participants for next-generation high-performance power supply applications. However, the threshold voltage ( V _th ) instability caused by short-circuit events jeopardizes the overall reliability of p-GaN gate HEMT and the power system in real applications. Hence, a noninvasive condition-based monitoring technique for critical device parameters is urgently required to enhance system safety without affecting the features of power electronic systems. In this paper, the threshold voltage instability dynamics of p-GaN gate HEMT under repetitive short stress were investigated to achieve high estimation accuracy and good monitoring efficiency. A double-phase adaptive neural network to predict the V _th degradation kinetics based on the historical degradation recordings of the target device is developed. The degradation process contains a monotonous increasing process and an oscillation process divided by random changing point subject to Weibull distribution. Based on such characteristics, the extreme learning machines were combined with classic activation functions and periodic activation functions to predict the threshold voltage tendencies of p-GaN HEMTs under repetitive SC stress. The experiment results validate that the developed model based on static investigations can provide degradation predictions with high accuracies. Besides, the proposed method endows substantial benefits for the conditional-based monitoring problem of other newly emerging semiconductor devices that feature multiple scenarios during the dynamic process and differences between individual units.