A Robust Data-Driven Method for Open-Circuit Fault Diagnosis of Power Switches in Three-Phase Inverters With Low-Quality Data

Machine learning (ML) techniques have shown great potential for power converter fault diagnosis. However, the data measured by the diagnostic processor may be corrupted in real-world applications, which would degrade the performance of ML-based diagnostic models. This article proposed a robust data-driven method for power switch open-circuit fault diagnosis under low-quality data issues with missing values, outliers, noises. At offline stage, a robust subspace matrix is first trained and adopted to recover the corrupted data from missing data and outliers. Then, the recovered data is further denoized through joint sparse coding and transform learning, where a transform weight matrix can be obtained. By using the processed data as the input, a random vector functional link network is trained to generate the diagnostic model. At online stage, real-time current signals are firstly recovered/denoized based on the trained matrixes and then sent to the trained diagnostic model to generate the diagnostic result. Simulation and real-time tests have demonstrated the high accuracy and strong robustness of the proposed method under various scenarios.