AI-Based Diagnosis Method for Multiple-Switch Faults in Three-Phase Inverters With Limited Labeled Samples

This paper proposes a novel fault diagnosis method that integrates signal processing and artificial intelligence (AI)-based techniques to detect open-circuit faults in multiple-switch three-phase inverters, achieving accurate fault detection with reduced labeled fault-scenario requirements for training. The increasing number of potential faulty switches complicates the process due to the growing data demands. To address this issue, a framework with multi-task learning (MTL) algorithm and one-phase current data are used during model training to reduce both the network size and training data requirements. For online fault diagnosis, the proposed method employs statistical feature-based signal processing of three-phase currents to identify faulty phases. The phase current data from the identified phases are processed using a three-parallel MTL algorithm for fault classification. Compared with existing methods, the proposed method improves fault detection accuracy, minimizes the training data requirements, and reduces the computational complexity. Its effectiveness has been validated through experiments, achieving a fault diagnosis accuracy of 99.85% across various conditions. The method requires less than 1.5 fundamental cycles of computational time and uses only three labeled fault scenarios out of 41 possible fault cases for training data acquisition.