LitePowerCD: A lightweight anomalous change detection model for power transmission and transformation scenarios

Abstract

Efficient and reliable detection of anomalous changes in electrical equipment is crucial for power grid safety. To address this, the LitePowerCD model is proposed for anomalous change detection(ACD) in power scenarios, using a siamese network with a lightweight EfficientNet_B2 backbone to efficiently extract dual-phase power scene features. Next, an efficient feature fusion module, Shallow Feature Fusion Module (SFFM), is designed. By introducing dilated convolutions and the Squeeze-and-Excitation module, the receptive field is expanded, and channel selection is enhanced, improving detection performance on small-scale change regions. Then, feature reconstruction is achieved through upsampling and skip connection, which retain and strengthen anomalous edge details and regional information. Moreover, a pixel-level classifier is improved to make more precise judgments, thereby reducing the false positive rate. Furthermore, a dataset of substation equipment images, covering various scenes, seasons, and lighting conditions, is constructed to support power equipment ACD research. Finally, experiments are conducted on the proposed model. Experimental results show that the proposed approach surpasses the previous state-of-the-art model by 1.07% on the F1-score, achieving 93.21%, while reducing the model size to only 0.23 MB (1/16 of prior methods) and significantly lowering computational cost, demonstrating a superior balance between efficiency and accuracy.