Research on multi-entity interaction oriented distribution grid operation scene understanding and intelligent recognition methods

In the field operations of distribution networks, the complexity of work scenes and the visual similarity of targets pose significant challenges to existing intelligent safety monitoring systems, leading to difficulties in target recognition and high false detection rates. These issues substantially hinder the safety assurance of power maintenance tasks. To address these challenges, this paper proposes an enhanced YOLOv8-based multi-object detection and scene understanding method tailored for distribution network operations. The proposed approach integrates a channel reconfiguration attention mechanism into the YOLOv8 framework to improve inter-channel feature interaction and enhance the saliency of key target regions. Additionally, a residual skip-connected multi-scale feature fusion structure (Res-PANet) is introduced to effectively integrate features across different scales, improving the detection of small objects in complex scenarios. A Swin-Transformer module is incorporated into the detection head to strengthen global perceptual capability and robustness against occlusion. Furthermore, a spatial relationship graph is constructed among detected objects, and a graph neural network is employed to model and extract topological interaction features. Experimental results demonstrate that the proposed method significantly enhances the accuracy of recognizing multi-object relationships and interactive behaviors in distribution network operation scenes, showing strong stability and promising application potential.