Transforming transformer diagnostics: Zero-shot learning powered by symptom descriptions and information fusion

Power transformer fault diagnosis is essential for maintaining grid stability, yet conventional approaches face challenges when diagnosing previously unseen fault types. This paper proposes a novel symptom-based information fusion framework for zero-shot diagnosis (SIFZ) to address this limitation. Unlike traditional methods, SIFZ does not rely on extensive labeled data or predefined fault categories. It employs a multi-layer, heterogeneous feature extraction process that simultaneously captures spatial, temporal, and symptom-related characteristics of fault data. The method leverages fault symptom descriptions to support cross-domain knowledge transfer and enables recognition of novel fault types by mapping observed symptoms to underlying fault conditions. Through this approach, SIFZ achieves robust fault identification without prior exposure to specific fault examples. Experimental results indicate improvements in diagnostic accuracy and robustness, with SIFZ achieving an average accuracy of 87.91% across different zero-shot diagnosis tasks, demonstrating the method's potential as a scalable and adaptable solution in dynamic environments.