WANG Peng, ZHOU Hualiang, ZHAO Chunhui, SU Zhantao, WANG Jing
Control Engineering of China. 2026, 33(4): 729-737.
To enhance the safety and operational intelligence of substations, automatic anomaly detection methods based on robotic inspection videos are widely adopted. However, achieving high-precision detection remains challenging due to the diversity of anomaly types, the complexity of anomaly features, and the scarcity of anomaly samples in substation equipment scenarios. To address this issue, a few-sample anomaly detection method for the substation equipment based on spatio-temporal feature fusion is proposed. Firstly, the LiteFlowNet is used to extract motion features from the template image and the comparison image, generating bidirectional optical flow maps with strong anomaly representation capability. Then, a dual-stream siamese network (DS-SiameseNet) is designed to integrate dual-modal features from the template image, the comparison image, and the bidirectional optical flow maps, and a dual-attention network is further introduced to refine multi-scale fusion features for accurate anomaly localization. The inspection dataset of a substation equipment is used to test the proposed method in the experiment. The experimental results show that, compared with the existing methods using image pair stitching with non-siamese networks, the proposed method effectively improves the anomaly detection accuracy of the substation equipment in scenarios with few samples.
Key words: Substation patrol inspection; optical flow estimation; dual-stream siamese network
