Self-Supervised Representation Learning-Driven Few-Shot Learning for On-Site GIS Partial Discharge Condition Assessment

Deep learning algorithms have achieved high assessment accuracy for gas insulated substation (GIS) partial discharge (PD) condition when provided with complete and abundant training datasets. However, in practical applications, most GIS equipment operates under normal conditions, and disassembly to identify PD types upon detection is impractical. This results in a lack of sufficient labeled data for training. To address the challenge of reliable GIS PD condition assessment under the constraints of limited labeled samples, this paper proposes a novel self-supervised representation learning framework to guide few-shot learning for on-site GIS PD condition assessment. First, contrastive-based self-supervised representation learning is employed to extract intrinsic features from large amounts of unlabeled GIS PD data, enhancing the model's feature representation capability. Subsequently, a dual-similarity metric learning approach is introduced to transfer the learned intrinsic features from unlabeled data to achieve generalized and robust evaluations of GIS PD conditions with few-shot scenario, effectively mitigating the issue of limited labeled data. Experimental results demonstrate that the proposed method achieves diagnostic accuracies of 95.36% for GIS PD type identification and 93.24% for severity assessment under practical on-site conditions. This work establishes a solid foundation for feature representation learning in the absence of labeled samples and promotes few-shot learning as a robust method for real-world applications.