Lightweight multi-class defect detection for intelligent substation inspection

Abstract

Accurate and efficient defect detection is essential for intelligent substation inspection, yet existing lightweight detectors often struggle to balance detection accuracy, computational complexity, and deployment efficiency while covering diverse inspection targets. This paper proposes a lightweight multi-perception object detection network based on YOLOv8 for practical substation inspection. The proposed method integrates a lightweight CFAT backbone with Partial Convolution, ADown downsampling, and Triplet Attention to improve feature extraction while reducing computational cost. A SlimNeck architecture is adopted for efficient multi-scale feature fusion, and a DyHead detection head further enhances object perception through dynamic attention. To better represent real inspection scenarios, a new dataset, Electric_Data, containing eight defect categories and seventeen typical inspection targets, is constructed. Experimental results demonstrate that the proposed method reduces model parameters by 33.2%, improves mAP@0.5 by 3.0%, and increases inference speed by 18.4% compared with YOLOv8s. Deployment on the NVIDIA Jetson Xavier NX further achieves a 46.8% improvement in real-time inference speed, indicating that the proposed framework is well suited for lightweight edge-based intelligent inspection systems.