Existing deep learning models struggle to balance accuracy and efficiency in segmenting fault areas within infrared images of substation equipment,which typically feature irregular shapes,small targets,and blurred edges.To address this, we propose a lightweight segmentation model based on an improved YOLOv8 framework.For model lightweighting,MobileNetV3 serves as the backbone, significantly reducing parameters and computational complexity through depthwise separable convolutions combined with an attention mechanism,while maintaining competent feature extraction.To enhance multi-scale irregular target perception,a Semantic Detail Injection Module is introduced, enabling effective fusion of high-level semantics with low-level detail features.Additionally,an Upsampling Inverted Bottleneck Fusion Structure is developed to address inherent infrared edge blurring,strengthening boundary delineation of fault areas.A composite loss function integrating Focal Loss and Dice Loss is employed to alleviate class imbalance and enhance robustness in complex backgrounds.Experimental results show the proposed method improves mAP by 2.38% over original YOLOv8 while reducing parameters by 26.46%.Benchmarking against four mainstream segmentation algorithms confirms its superior accuracy and computational efficiency, demonstrating strong potential for intelligent power equipment maintenance applications.

Insulators are often significantly impacted by target occlusion in complex environments during UAV inspections. To address this issue, an insulator fault detection algorithm based on YOLOv5 is proposed, into which an attention mechanism and cross-scale adaptive weighted feature fusion are incorporated. To suppress interference from complex backgrounds, a Coordinate Attention (CA) module is added to the backbone network. Additionally, the original FPN + PAN structure is replaced with a BiFPN to achieve more effective multi-scale feature fusion. To tackle false detections and missed detections caused by target occlusion, an adaptive weighted feature fusion module is introduced. Moreover, a Power IoU loss with an tuned power factor is used to improve bounding box regression. Furthermore, ghost convolution is integrated into the YOLOv5 backbone to improve detection speed. To verify the algorithm’s effectiveness, training and test samples were prepared using data collected by a power department over three years of UAV inspections. The proposed algorithm was compared with three mainstream algorithms. Results show that our algorithm accurately detects insulator faults under various complex conditions with excellent real-time performance.