Automatic Detection Algorithm of Typical Hidden Dangers in Substation Based on Dynamic Convolutional Feature Extraction

Abstract

With the advancement of smart grid construction, substation equipment, being perpetually exposed to complex environments, is prone to latent hazards such as meter malfunctions and insulator cracks. Additionally, personnel violations may trigger safety incidents, directly jeopardising grid reliability. However, traditional manual inspections suffer from low efficiency and high rates of missed defects, while existing detection methods struggle to accommodate irregular defect classifications and multi-type defect characteristics, failing to meet engineering demands for real-time response and precise identification. To address these challenges, this study proposes a substation hazard detection framework (YOLOv10_DSE) based on an enhanced YOLOv10 (You Only Look Once version 10) architecture, designed to tackle multi-type hazard detection within complex substation scenarios. Firstly, a dynamic feature extraction module (C2fDSC) was designed, employing dynamic snake convolutions to enhance adaptive sampling capabilities for small targets and irregular defects. Secondly, a self-integrated attention module head (SEAMHead) was introduced to decouple localisation and classification tasks, thereby improving multi-type hazard discrimination accuracy. Finally, a bounding box regression loss function (inner_CIoU) was adopted to optimise small target localisation and irregular shape fitting. Experiments demonstrate that on a substation dataset containing 17 defect types, this method achieves mAP@0.5 and mAP@0.95 of 73.3% and 48.2%, respectively, representing improvements of 2.6% and 1.6% over the YOLOv10 baseline. This provides an efficient and reliable solution for multi-type defect detection in substations, holding significant engineering value for ensuring the secure operation of power grids.OPEN ACCESS Received: 15/10/2025 Accepted: 17/12/2025

Document