Failure Prediction of Power Electronic Devices

Sammanfattning

Power electronic devices have become integral components in modern consumer and transportation industries. Predicting the failure or health status of these devices not only ensures operational safety and prevents catastrophic consequences but also leads to reduced downtime and operational costs. However, failure or health status prediction represents a complex problem marked by numerous intrinsic and extrinsic variables, leading to different lifetimes of devices. Additionally, selecting relevant precursor signals that effectively capture the underlying failure mechanisms and overcoming time-series prediction challenges, such as handling dynamic and non-linear behaviors, are crucial for accurate predictions. In the thesis, three models—Kalman filter (KF), Particle filter (PF), and Autoregressive Integrated Moving Average (ARIMA)—are applied, compared, and evaluated for failure or health status prediction of power electronic devices using Power Cycling (PC) test data for power diodes. Among the models, the KF demonstrates the most significant performance while consuming the least amount of time. The PF achieves the second-best performance and the third-best time consumption. Meanwhile, the in-sample ARIMA model delivers the third-best performance and the second-best time consumption. Finally, the out-of-sample ARIMA model ranked the lowest in both performance and time consumption. These results suggest that dynamic models, specifically the KF and PF, exhibit superior generalization capabilities across different devices. This underscores the potential of dynamic models for enhancing predictive accuracy while optimizing computational efficiency in the context of real-time power electronic device health monitoring.