Identification of IGBT Bonding Wire Degradation Based on Feature Extraction and State Classification

As a vulnerable part of insulated gate bipolar transistor (IGBT), condition monitoring of bonding wires can detect potential faults at an early stage, thus avoiding threats to system safety. In this paper, we propose a monitoring method that combines feature extraction and condition classification, using multi-cycle and low-frequency operation data to construct a high-dimensional matrix, analyzing spectral properties and constructing linear eigenvalue statistics based on random matrix theory to extract potential information. An optimized random forest model using cross validation and particle swarm optimization algorithms ensures that the classification task benefits from the selected features and the optimized model. The results show that the feature extraction method can be highly adaptable to high-frequency action switches and is able to identify latent patterns. The optimized random forest model accurately classifies early bond wire breakage states, requires a lower sampling frequency than traditional methods, and still performs well in unbalanced data.