Anomaly Detection Method for Converter Transformer Oil Conservators Using Endoscopic Visual Perception

The metallic shell of converter transformer oil conservators prevents direct visual monitoring of the internal capsule, creating a long-standing monitoring blind spot that may leave hidden defects such as top gas accumulation undetected. To address this issue, this paper proposes an endoscopic video monitoring device installed inside the oil conservator capsule, enabling continuous visual monitoring without compromising insulation performance. On this basis, an anomaly detection method is established: region-of-interest (ROI) templates are defined in areas where the capsule normally adheres to the metallic inner wall of the oil conservator; normalized cross-correlation (NCC) is applied for template matching, and kernel density estimation (KDE) statistically analyzes the distribution of matching scores from normal samples to determine per-template thresholds. For valid template localizations, Euclidean distances among template centroids are extracted as geometric features and fed into an Isolation Forest for unsupervised anomaly scoring. Validation on field data from a 500 kV converter station demonstrates that the method achieves a precision of 98.78%, a recall of 96.43%, and an F1-score of 97.59%. The results indicate that the proposed method can distinguish normal capsule breathing from capsule morphological abnormalities caused by top gas accumulation, providing a basis for online monitoring of oil conservator capsules.