Abstract: Due to the complexity and differences of defect types in substations, traditional deep learning models for defects detection lack comprehensive response ability. It proposes a sematic distance based decoupling detection model. Firstly, the decoupled model structure is determined by clustering defect classes according to the semantic information distance between each other. Then, the weighted anchor fusion and local prediction loss techniques are used to improve the model performance. Meanwhile, the decoupled non-maximum suppression strategy is proposed to accelerate the model inference process. The experiment results show that the mean average precision of the model reaches 69.68%. Compared with YOLOX, which has been recognized as the best real-time object detection model, the accuracy of proposed model is improved by 1.36 percentage points, the parameter quantity is reduced by 5%, and the inference speed is improved by 34%.