In order to solve the problems of complex and changeable background, large intra-class gap, drastic scale change and semantic bias of substation defects, an improved YOLOV9 algorithm is proposed. Firstly, the dynamic convolution module (ODCONV) is introduced into the YOLOV9 network structure, and the dynamic convolution module is used to improve the adaptability of the network to complex and changeable backgrounds and large gaps within classes. Secondly, Convolutional Block Attention Module (CBAM) is introduced for optimization, which significantly reduces the semantic bias of substation defect detection. Finally, in view of the drastic change of the target scale in the data set, Inner-IOU is introduced to improve the sensitivity of the network to the scale, accelerate the network convergence, and improve the detection effect of the network on substation defects. In order to adapt to the task of substation defect detection, a special substation defect dataset is constructed in this paper. The experimental results show that the improved YOLOV9 algorithm achieves good performance in the substation defect identification task. Compared with the original YOLOV9 algorithm, the improved algorithm presents stronger capabilities in the identification of complex substation defects and small-sample targets. The proposed algorithm improves mAP50 by 1.45% and mAP by 1.13% on our dataset, achieving state-of-the-art object detection algorithms.