A Dynamic Physics-Guided Ensemble Model for Non-Intrusive Bond Wire Health Monitoring in IGBTs

TL;DR

This paper introduces a new physics-guided machine learning model to monitor bond wire health in IGBTs, improving accuracy and reliability in power electronics.

Contribution

The novel contribution is a physics-constrained ensemble learning framework that integrates multi-physical features and adaptive model fusion for bond wire health monitoring.

Findings

The proposed framework achieves a mean absolute error of 0.0066 V and R2 of 0.9998 in predicting Vce-on.

The model shows a 48.4% improvement over individual base models while maintaining 99.1% compliance with physical constraints.

The method harmonizes data-driven learning with physical principles for robust health monitoring in power electronics.

Abstract

Bond wire degradation represents the predominant failure mechanism in IGBT modules, accounting for approximately 70% of power converter failures and posing significant reliability challenges in modern power electronic systems. Existing monitoring techniques face inherent trade-offs between measurement accuracy, implementation complexity, and electromagnetic compatibility. This paper proposes a physics-constrained ensemble learning framework for non-intrusive bond wire health assessment via Vce-on prediction. The methodological innovation lies in the synergistic integration of multidimensional feature engineering, adaptive ensemble fusion, and domain-informed regularization. A comprehensive 16-dimensional feature vector is constructed from multi-physical measurements, including electrical, thermal, and aging parameters, with novel interaction terms explicitly modeling electro-thermal stress coupling. A dynamic weighting mechanism then adaptively fuses three specialized gradient boosting models (CatBoost for high-current, LightGBM for thermal-stress, and XGBoost for late-life conditions) based on context-aware performance assessment. Finally, the meta-learner incorporates a physics-based regularization term that enforces fundamental semiconductor properties, ensuring thermodynamic consistency. Experimental validation demonstrates that the proposed framework achieves a mean absolute error of 0.0066 V and R2 of 0.9998 in predicting Vce-on, representing a 48.4% improvement over individual base models while maintaining 99.1% physical constraint compliance. These results establish a paradigm-shifting approach that harmonizes data-driven learning with physical principles, enabling accurate, robust, and practical health monitoring for next-generation power electronic systems.

Figures