Abstract:

The condition of coal mine power equipment condition the power supply safety of mine shafts. UAV-borne infrared thermal inspection has become an important monitoring method. However, existing lightweight detection algorithms lack satisfactory performance in identifying early tiny defects under the conditions of low infrared contrast, large target scale differences and complex thermal backgrounds, making it difficult to balance real-time performance and detection accuracy for edge devices. To address these problems, MobileNetV3 lightweight network was adopted as the backbone, and a multi-kernel dynamic convolution mechanism was embedded into inverted bottleneck residual units. Multiple convolution kernels ware fused through input-aware attention weights to adaptively adjust the convolution operators. Parameter-sensitivity experiments confirmed that setting the kernel number to 4 achieved the optimal balance between detection performance and computational cost. Meanwhile, the Efficient Channel Attention (ECA) module was introduced into shallow layers of the backbone network to highlight thermal edge details and suppress background thermal noise. The Squeeze-and-Excitation (SE) module was retained in deep layers to strengthen high-level semantic discrimination capacity. The Ghost module was integrated in the channel expansion stage to cheaply expand feature maps via linear transformation. The three modules formed a collaborative link of “denoising and retaining details in shallow layers, enhancing semantics in deep layers, and supplementing textures during channel expansion”, enriching feature representation with minimal extra computational overhead. Experiments were conducted on a self-built infrared image dataset of coal mine power equipment. Containing 5 000 images of five typical equipment types and their defects: disconnecting switches, insulator strings, transformer bushings, cable joints and surge arresters. Experimental results showed that the improved model achieved an accuracy of 88.5%, an mAP of 86.7% at IoU threshold 0.5 and an F1-score of 86.6%. The computational complexity of the model was around 3.18 GFLOPs, and the inference latency on Central Processing Unit (CPU) was approximately 39 ms. Ablation experiments confirmed the collaborative gain of dynamic convolution and hierarchical feature enhancement, with optimal performance obtained at the kernel number K=4. Compared with recent mainstream lightweight detection algorithms, including YOLOv8n, YOLOv9-Tiny, RT-DETR-R18 and YOLOv10-N, the proposed method achieves the optimal comprehensive balance between accuracy and efficiency. Visual detection results also demonstrated that the model maintained high detection confidence for all types of power equipment under low resolution, heavy noise and occlusion. The proposed method greatly improved infrared defect detection performance while meeting the real-time reasoning requirements for UAV edge terminals, demonstrating engineering practicality in complex underground coal mine scenarios.

Key words: coal-mine power equipment, lightweight object detection, infrared inspection