Lightweight Machine-Learning-Based Diagnosis for Power Electronic Systems Subject to Imbalanced Data

• Published in: IEEE journal of emerging and selected topics in industrial electronics (Print) Vol. 5; no. 2; pp. 733 - 744
• Main Authors: Deng, Qingli, Feng, Xiaoyun, Zhao, Shuai, Gou, Bin, Wang, Huai, Ge, Xinglai
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.04.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Data models; Electronic systems; Ensemble learning; Fault diagnosis; Feature extraction; hybrid feature selection scheme; imbalanced data; Machine learning; Oversampling; Power electronics; Pulse width modulation; Real-time systems; Rectifiers; safe-level synthetic minority oversampling technique (safe-level SMOTE)
• ISSN: 2687-9735; 2687-9743
• DOI: 10.1109/JESTIE.2024.3358729

Machine-learning-based fault diagnosis methods have gained extensive applications in power electronic systems. One of the key challenges for field implementation is the rarity of fault history data in practical systems. It leads to an imbalanced ratio between normal and fault state data, causing high accuracy for the majority but severely reducing that of the minority class. To address this issue, a method integrating of oversampling and ensemble learning is proposed in this article. First, the feature extraction and hybrid selection schemes are designed to obtain the most critical features, thereby enhancing the performance of the sampling process. The safe-level synthetic minority oversampling technique is then applied to sample the minority classes to balance the historic database. An ensemble random vector functional link network learning model, which is computationally light, is developed as the diagnostic model to further improve the accuracy. The method is demonstrated by an experimental study on the sensor and open-circuit faults of single-phase rectifiers.