A Distance-Based Health Indicator and Its Use in an Interacting Multiple Model for Failure Prognosis in Power Electronic Devices

• Published in: IEEE transactions on reliability pp. 1 - 15
• Main Authors: Yang, Qian, Joshi, Shailesh, Viviano, Raymond, Ukegawa, Hiroshi, Pattipati, Krishna R.
• Format: Journal Article
• Language: English
• Published: IEEE 04.02.2025
• Subjects: Adaptation models; Data models; Degradation; Extended Kalman filter (EKF); interacting multiple model (IMM); K-means; Mathematical models; MOSFET; power electronics; Predictive models; remaining useful life (RUL); Self-organizing feature maps; self-organizing map (SOM); Semiconductor device modeling; Switches; Vectors
• DOI: 10.1109/TR.2025.3526594
• ISSN: 0018-9529

Power electronic (PE) reliability is critical to electric vehicle performance and safety. Thus, it is vital to predict the remaining useful life (RUL) of components that are subject to predictable degradation. Here, we propose a RUL estimation framework for PE components. The framework has two consecutive phases: Generation of distance-based health indicators through an unsupervised learning procedure, such as self-organizing map (SOM) or K-means clustering, and subsequent deployment of interacting multiple model (IMM) that integrate linear and extended Kalman filters with varied degradation profiles to forecast future values of the indicator and RUL. Specifically, a nominal SOM or K-means model is learned, using the on -state median signal data from the PE component. The indicator is then calculated by measuring the distance between the test vector and the cluster center. To adaptively track the health indicator and its rate of change, accounting for the noise intrinsic to degradation processes, various degradation profiles, and the measurement system, the IMMs are applied. The RUL is evaluated as the difference between a predefined threshold and the health indicator estimate, divided by the present degradation rate. Validation of the framework involved accelerated aging experimental datasets, encompassing both low-frequency and high-frequency switching scenarios. The results reveal the framework's versatility and potential for implementation across diverse applications.