An improved YOLOv9-based method for identifying insulator temperature anomalies

Abstract Abnormal heating of transmission-line insulators is often an early indicator of impending faults. Timely and accurate identification of overheated regions in infrared imagery is therefore critical for the safe operation of power grids. However, current inspection workflows still rely heavily on manual interpretation. They are vulnerable to missed detections and misjudgments under complex thermal backgrounds and large image volumes. In addition, deployment on UAVs and other edge platforms is constrained by computation and power budgets. These constraints make it difficult for general-purpose detectors to achieve both high accuracy and real-time performance in practice. Here we propose IODNet, an edge-oriented method for detecting insulator overheating anomalies in infrared images. Built on YOLOv9, IODNet replaces the backbone with ShuffleNetv2 to reduce computational cost. It further introduces an MSBlock module in the feature-fusion stage to strengthen representations for tiny targets and multi-scale hot spots. For optimization, we adopt the Inner-WIoU loss to improve localization robustness and generalization. Experiments show that IODNet reduces computational complexity and improves inference efficiency while maintaining high detection accuracy, making it well suited to resource-constrained engineering deployments.