Zero-Shot Learning for Unknown Fault Diagnosis in High-Voltage Circuit Breakers via Envelope Spectrum Semantic Construction and Embedding

The reliable diagnosis of mechanical faults in high-voltage circuit breakers (HVCBs) is crucial for ensuring the safe and stable operation of power systems. However, existing diagnostic methods for HVCBs are typically limited to known fault types within the training dataset, making them inadequate for handling scenarios involving unknown fault classes and compound faults that emerge during operation. To address these challenges, we propose a novel zero-shot learning (ZSL) based on envelope spectrum semantic construction and semantic embedding space for diagnosing unknown and compound faults in HVCBs. First, a deep learning model with weighted cross-entropy loss is utilized to extract discriminative features of known single faults under imbalanced data distribution. Subsequently, ZSL is employed to encode single fault attributes into a semantic space, enabling the representation of compound fault semantics. Considering that the envelope spectrum effectively characterizes fault features in HVCBs, fault signals are preprocessed using the envelope spectrum to enhance feature representation. The physical meaning of the envelope spectrum is further leveraged to construct semantic representations for both single and compound faults. Finally, a fully connected neural network maps the sample features into the semantic space, where similarity measurements are used to identify unknown and compound faults. The proposed method is validated on an HVCB dataset, demonstrating an average recognition accuracy of 96.63% for unknown faults and 92.45% for compound faults under imbalanced data distribution, even in the absence of compound samples in the training set. These results confirm the effectiveness and generalization capability of the proposed approach.