Abstract

In the process of substation anomaly detection, abnormal samples of substation equipment are scarce. Anomaly detection methods such as object-based detection are difficult to effectively train models for all equipment and are unable to deal with unknown anomalies, resulting in poor robustness. To solve this problem, a weakly supervised learning method is used to detect anomalies through the distance between the image to be measured and the image of normal equipment, so that it can be applied to all devices. Substation anomaly detection data set is constructed, and image pairs are used as training samples to expand the number of negative samples. A lightweight substation change detection network (TSCDNet+) is proposed, which uses a Siamese network structure to add an attention module to reduce the interference of the background environment. The full connection layer was used to calculate the distance of feature vectors to enlarge the feature gap between positive and negative samples. The multi-scale information fusion structure was added, and the feature map was used to replace the feature vector to calculate the distance. The experimental results show that TSCDNET+ has good robustness under different thresholds, and the F1 score of the model with the best effect is 0.945. As it does not rely on the category characteristic information of equipment and anomaly, this method has a good universality, can meet the detection of an unknown anomaly, and still has a good detection effect for new equipment that has not been trained.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call