Improved IGBT Aging Diagnosis for Three-Phase Inverters via Phase-Angle Feature Redesign and Kernel SHAP Analysis

This study proposes an improved artificial neural network (ANN)-based framework for diagnosing the aging state and predicting the remaining useful life (RUL) of Insulated Gate Bipolar Transistor (IGBT) modules in three-phase inverter systems. Building upon a prior 15,625-scenario simulation dataset, the proposed approach fundamentally redesigns the input feature space by replacing the three-phase voltage rates of change with phase angles of symmetrical components ( $\angle V_{0}, \angle V_{1}, \angle V_{2}$ ), eliminating redundant information and enabling accurate estimation of the inverter aging severity. Two independent ANN models—a classification model for risk-level identification and a regression model for continuous aging index estimation—share the same 6-dimensional input space derived from symmetrical component transformation. The classification model achieves 96.86% accuracy under noisy conditions (SNR 20 dB), and the regression model attains $R^{2}=0.9028$ , both representing improvements over the prior study under their respective diagnostic frameworks. Kernel SHAP global sensitivity analysis, cross-validated against an ablation ground truth, shows that gradient-based attribution fails to capture the positive-sequence ( $V_{1}$ ) contribution (0.00% gradient vs. 18.5%/41.1% SHAP) under multicollinearity among inputs ( $r=0.853$ between $V_{0}$ and $V_{2}$ ). A 4-input ablation experiment empirically validates the necessity of all six features, with classification accuracy dropping to 35–89% and regression $R^{2}$ collapsing to −8.10 upon $V_{1}$ removal. The proposed software-based solution requires no additional hardware and is broadly applicable to predictive maintenance in electric vehicles, drones, and renewable energy systems.