Detection of Wire Lift-Off in Si-IGBTs and SiC-MOSFETs Using Machine Learning on Switching Waveforms

Power cycling degradation in power modules is one of the most common failure modes in power electronic systems. This study presents a method to detect wire lift-off using machine learning analysis of switching waveforms of gate voltage and emitter/source current in Si-IGBT and SiC-MOSFET power modules. Five key aspects are investigated as follows: (1) evaluation of different detection signals and their combinations, (2) segment-wise analysis to identify waveform regions that contribute most to classification, (3) impact of waveform sampling rate on detection accuracy, (4) validation using mixed datasets from multiple modules, and (5) comparison between Si-IGBT and SiC-MOSFET devices. The turn-off gate voltage waveforms provide the highest accuracy, and a high accuracy of about 99 % was achieved even with downsampling from 2.5 GS/s to 100 MS/s. The mixed datasets show no degradation in accuracy, and the optimum gate drive condition for high accuracy differs between Si-IGBTs and SiC-MOSFETs. Overall, the proposed method achieves over 99 % classification accuracy, demonstrating its potential for reliable and low-cost monitoring of power device degradation.