Anomaly Detection of Partial Discharge Signals in Transformers Based on Self-Supervised Temporal Contrastive Learning

Partial discharge (PD) identification in transformers is crucial for ensuring operational safety through timely fault detection. This study introduces the Self-supervised Temporal Anomaly Detection via Contrastive Learning (STAD-CL) model to detect temporal anomalies in time-series sensor data. The approach employs an end-to-end framework for feature extraction, utilizing wavelet transform convolution-1d (WTConv-1d) to capture multi-frequency features by combining convolutional operations with wavelet transform. Additionally, dilated convolution-1d (DConv-1d) is introduced to extract temporal features across varying scales via dilated convolution. The temporal slicing strategy targets local features within the signal, thereby reducing redundancy. And the positive sample selection learns position-invariant representations from overlapping regions, effectively mitigating representation collapse. A novel dual-level contrastive loss, incorporating instance-level and temporal-level contrasts with an attention mechanism, further enhances learning performance. Finally, an anomaly detection criterion is established, effectively identifying abnormal signals by combining segment-wise signal scores with predefined thresholds. Experimental results demonstrate that STAD-CL achieves state-of-the-art results on benchmark datasets for transformer PD monitoring.