PG-YOLO: a YOLOv12-based detector for multi-class anomaly detection in power grid maintenance

Abstract

Power grid maintenance involves diverse abnormal targets, such as unsafe worker behavior, abnormal meter readings, cabinet-door faults, foreign objects, oil leakage, and equipment damage. These targets often appear with scale variation, occlusion, and complex backgrounds, which challenges real-time inspection systems. This paper proposes PG-YOLO, a lightweight detector based on YOLOv12 for multi-class anomaly detection in power grid maintenance. A dataset with 28 categories and 188 353 annotated instances is built to support fine-grained recognition of normal and abnormal states. In PG-YOLO, wConv2D is introduced to reduce redundant computation and enhance local responses. A2C2F_CGLU is used to strengthen spatial–channel feature interaction, and SPPF&C2PSA is added to improve high-level semantic representation. Experiments on the constructed dataset show that PG-YOLO achieves 83.7 precision, 73.9 recall, 79.1 mAP 50, and 56.3 mAP 50:95 with 2.98 M parameters, 5.0 GFLOPs, and 227.28 FPS. The ablation study verifies the contribution of each module, and Grad-CAM visualizations show more compact activation on anomaly-related regions.