Generative Zero-Shot Learning for Partial Discharge Diagnosis in Gas-Insulated Switchgear

Abstract

Class imbalance exists widely in partial discharge (PD) diagnosis of gas-insulated switchgear (GIS). On the one hand, while atypical PDs occur sometimes but the sample is scarce. On the other hand, due to the contingency and concurrency of PDs, the demand for multi-source PD increases exponentially, making it difficult to obtain sufficient data. This study focuses on zero-shot diagnosis in extreme cases where only typical defects are available and where atypical and multi-source PDs are not available during training. And we propose a semantic rectifying discriminative generative adversarial network (SRDGAN) for zero-shot diagnosis. The proposed SRDGAN trains the generator from the visible samples and semantic attributes, and generates the unseen samples from the unseen semantic attributes and trains the classifier for zero-shot diagnosis. First, a semantic rectifying module is designed to correct the structure between the visual and semantic space to make the semantic features distinguishable. Then, the latent discriminative attributes are extracted from the visual features, and an attribute embedding module is designed. Finally, PD diagnosis is performed on feature generation and classification modules. The proposed SRDGAN is validated on two datasets. The experimental results illustrate that the SRDGAN solves zero-sample GIS PD diagnosis with >90% accuracy.