A substation anomaly detection method based on improved Siamese networks

Abstract

An urgent problem in the process of substation anomaly detection is the scantiness of abnormal substation equipment samples. Conventional anomaly detection methods which use object-based detection are difficult to distinguish every type of equipment and unknown anomalies, leading to poor generalization. A substation anomaly detection method based on lightweight Siamese network is proposed to solve these problems, the distance between the image to be measured and the image of normal will be calculated in the method for detecting anomalies, so our method is able to deal with unknown samples, and data enhancement methods such as rotation, mirroring, and symmetry are used to expand the number of negative samples. The improved Siamese network adds ECA attention module to decrease the interference of background environment, ECA module performs channel feature learning on the compressed feature map, multiplies the original input feature map and the channel feature maps, and finally outputs the feature map with channel attention. The full connection layer calculates the distance of feature vector instead of the conventional Euclidean distance or cosine distance, expanding the difference value between positive and negative samples, and clarifying the threshold selection. This method has a good universality, it can meet the detection of various background environment and has high accuracy of detection work.