An Online Convolutional Neural Network Based Method for Open-Circuit Fault Diagnosis in Three-Phase Inverters Under Extremely Unbalanced Loading Condition

Diagnosing open-circuit (OC) faults in three-phase inverters becomes particularly challenging under extremely unbalanced loading conditions. In such cases, the phase current with a greatly reduced amplitude renders the predefined thresholds in rule-based methods unreliable and is tend to be neglected in feature extraction in data-driven methods, thereby reducing overall diagnosis accuracy. In addition, since existing data-driven methods require at least one complete fundamental cycle of three-phase current for analysis, fault detection is inherently delayed by one fundamental cycle. To overcome these limitations, this paper proposes a convolutional neural network (CNN)–based OC fault diagnosis method with accuracy-driven feature extraction and classification. The method encodes the three-phase currents as a current- vector trajectory in the \alpha–\beta domain and converts each trajectory into a binary matrix representation. By leveraging the intrinsic feature extraction capability of CNNs and backpropagation-based training, feature extraction and classification are integrated within one network and are directly guided by diagnosis accuracy, preventing accuracy degradation under extremely unbalanced loading condition. Two CNNs are employed: the first CNN achieves sub-cycle fault detection by identifying trajectory distortions, while the second CNN recognizes the trajectory pattern under faulty conditions and pinpoints the specific faulty switches one fundamental period later. Experimental results demonstrate that the proposed method accurately classifies all 21 OC fault cases and the healthy condition, achieving 100% accuracy under both balanced operation and extremely unbalanced loading conditions with a phase-current unbalance rate (PCUR) of up to 88%.