Research on Substation Equipment Defect Detection Technology Based on an Improved RT-DETR Model

To address the challenges in intelligent inspection of substation equipment, including difficulties in acquiring defect image data, the high complexity of equipment, and the low efficiency of defect detection technologies, this paper conducts research on defect detection techniques for substation equipment based on an improved RT-DETR model. Additionally, a specialized dataset for substation equipment defect detection is constructed. First, the backbone of RT-DETR is replaced with the lightweight FasterNet, and deformable convolution (DCNv4) and channel attention (ECA) are introduced to achieve efficient feature extraction for multi-scale targets, improving detection accuracy while reducing the number of parameters. Moreover, an improved cascaded attention encoder is used to replace the original encoder, enabling the network to focus more on the target regions, further reducing computational redundancy and enhancing model efficiency. Finally, extensive comparative experiments are conducted on the self-constructed substation equipment defect dataset. The results show that the improved model achieved a 2.5% increase in average detection accuracy mAP@0.5 and a 2.2% increase in mAP@0.5:0.95 compared to the baseline model. Furthermore, the model’s parameters are reduced by 13.9%, GFLOPs decreased by 5.1%, and the frames processed per second (FPS) increased by 35%. Overall, the model outperforms other advanced defect detection models and is better suited to meet the demands of defect detection for multi-class and multi-scale targets in complex environments.