Research on the Diagnosis Method of Unseen New Faults and Composite Faults of High Voltage Circuit Breaker via Zero-Shot Learning

In recent years, data-driven methods have developed rapidly in fault diagnosis of high-voltage circuit breakers (HVCBs). However, in the face of unseen fault and compound faults that have no historical records in engineering practice, there are still shortcomings such as insufficient fault feature learning and high misdiagnosis rate. To address the above issues, we propose zero-shot learning for unknown classes and composite fault diagnosis in HVCB. First, this paper constructs a semantic attribute description that characterizes HVCB faults to obtain a vector representation of the fault description. Then, a deep attention residual convolutional network is constructed to extract discriminative features. Finally, an attribute learning network is constructed, which is trained by the characteristics of visible faults, and the attribute vectors of unseen fault samples are predicted by the attribute learning network to realize the diagnosis of unseen faults. Experimental results show that the proposed zero-shot learning achieves >90% diagnostic accuracy for unseen classes of new faults and compound faults, which is significantly better than other methods. It has laid a solid foundation for the diagnosis of unseen new faults and composite faults.