Performer-KAN-Based Failure Prediction for IGBT with BO-CEEMDAN.

Area of Science:

• Power Electronics
• Reliability Engineering
• Artificial Intelligence

Background:

• Insulated Gate Bipolar Transistors (IGBTs) are vital in power electronics but are prone to failure.
• Predicting IGBT failures is essential for ensuring system reliability and preventing catastrophic events.

Purpose of the Study:

• To develop an accurate and efficient prognostic model for IGBT failure prediction.
• To enhance real-time health monitoring and fault prediction capabilities for industrial applications.

Main Methods:

• Utilized Bayesian-optimized CEEMDAN (BO-CEEMDAN) for efficient extraction of fault features from IGBT data.
• Proposed a novel prognostic model, Performer-KAN, combining the Performer's FAVOR+ mechanism with Kolmogorov-Arnold Network (KAN) for enhanced nonlinear approximation.

Main Results:

• The Performer-KAN model demonstrated superior prediction accuracy, validated by metrics like MAE, RMSE, and R².
• Achieved higher predictive precision compared to six representative deep learning models.
• Maintained low computational overhead, indicating practical applicability.

Conclusions:

• The proposed Performer-KAN model offers a practical and effective solution for real-time IGBT health monitoring.
• The integration of BO-CEEMDAN and Performer-KAN significantly improves IGBT failure prediction accuracy and reliability.