Prediction of IGBT Gate Oxide Layer's Performance Degradation Based on MultiScaleFormer Network

Area of Science:

• Power Electronics
• Reliability Engineering
• Machine Learning

Background:

• Insulated gate bipolar transistors (IGBTs) are critical in power electronics, with gate oxide degradation being a key failure mode.
• Effective health management requires accurate prediction of IGBT performance degradation.

Purpose of the Study:

• To identify a reliable precursor for IGBT gate oxide degradation.
• To develop and validate an advanced degradation prediction model for IGBTs.

Main Methods:

• Gate leakage current was selected as the fault precursor parameter.
• Feature selection and fusion techniques (time domain analysis, grayscale correlation, Mahalanobis distance, Kalman filter) were employed.
• An improved degradation prediction model, MultiScaleFormer, inspired by iTransformer, was developed.

Main Results:

• A health indicator was successfully extracted to characterize IGBT gate oxide aging.
• MultiScaleFormer demonstrated superior fitting accuracy compared to LSTM, CNN, SVR, GPR, CNN-LSTM, and Transformer models.
• The MultiScaleFormer model achieved a mean absolute error (MAE) as low as 0.0087.

Conclusions:

• Gate leakage current is a feasible fault precursor for IGBT gate oxide failure.
• The MultiScaleFormer model offers a feasible and accurate approach for predicting IGBT performance degradation.