A Real-Time Edge Inference Method for Insulator Contamination Detection With YOLOv11-ssL

Insulators are a critical component of power transmission and distribution lines, requiring extensive in-service deployment. The accumulation of dust, bird droppings, rain, and snow leads to contaminated insulators, which can cause contamination flashovers, resulting in faults. Currently, manual contact-based detection of insulator contamination is costly and hazardous. In noncontact detection, image-based methods have developed rapidly, but the large data volume makes cloud transmission bandwidth-intensive and may result in high latency, which hinders real-time detection. This article aims to ensure the accuracy of contamination detection while reducing model complexity and parameter count, enabling low-cost real-time contamination inference on edge devices. The existing segmentation variant of YOLOv11 (YOLOv11n-seg) can already meet the accuracy requirements for insulator segmentation; therefore, this article proposes a lightweight variant of YOLOv11 (YOLOv11-ssL) network to reduce the model size. First, to accurately capture insulator features, this article designs a novel C3k2-star structure to enhance the information representation capability. Second, a lightweight shared convolutional separated batch normalization detection (LSCSBD) head structure is proposed to reduce the model size. Finally, layer-adaptive magnitude-based pruning (LAMP) is applied to YOLOv11-ssL to prune weights, significantly reducing the model size and computational workload. After segmenting the insulator, this article introduces the differential evolution and mutual information enhanced minimum redundancy maximum relevance algorithm (DEMIC-mRMR) algorithm to select the color features most correlated with the insulator contamination level; furthermore, Gaussian process regression (GPR), a nonparametric regression model, is used to predict the insulator contamination level effectively. Experimental results showed that YOLOv11-ssL reduces the computational load and the number of parameters by 44.1% and 82.8%, respectively, while maintaining a mean average precision (mAP) without substantial decrease, and achieves an inference frame rate of 1144.5. Meanwhile, $R^{2}$ of insulator contamination detection reached 0.8674, satisfying the requirements for real-time and accurate contamination detection on edge inspection devices.