CA-YOLO for abnormal behavior detection in power systems.

Abstract

Under the ”dual-carbon” objectives, the digital and intelligent transformation of power systems has accelerated, driving heightened demand for intelligent security. However, issues persist regarding insufficient detection accuracy for human abnormal behavior. To address this, this paper constructs a dataset covering climbing, vaulting,and normal behavior and proposes an abnormal behavior detection model CA-YOLO based on an improved YOLO11n architecture and Coordinate Attention(CA) specifically for this dataset. During the improvement process, the advantages of CA are fully utilised. Firstly, Receptive Field Coordinate Attention Convolution(RFCAConv) is introduced to enhance key region localisation capabilities. Secondly, CA is embedded into the backbone C3k2 module to strengthen multi-scale feature extraction. A CAFusion module is designed to integrate CA for optimising multi-level feature fusion. Additionally, an EMAM-MPDIoU loss function is developed to improve bounding box regression accuracy. Finally,the CA-YOLO model is validated using our self-built abnormal behavior dataset. Experiments demonstrate that compared to YOLO11n, the improved model achieves respective increases of 3%,7%,3.2% and 3% in mAP@0.5, mAP@0.5:0.95, precision (P) and recall (R), significantly enhancing detection accuracy and robustness. This work offers technical support for intelligent security applications in new-type power systems.