Abstract:  Aiming at the problems of large background interference in infrared images of substation equipment, various types of thermal defects, and the inefficiency of existing fault diagnosis methods, which are difficult to meet the actual inspection application requirements, a thermal defect identification and diagnosis method for substation equipment based on improved YOLO and Resnet is proposed. Firstly, construct a typical infrared image dataset of substation equipment, use convolutional kernel decomposition and multi-layer feature fusion technology to improve the YOLOv4-Tiny algorithm, locate the faulty equipment and obtain a priori frame of the equipment; Then, we propose a Res_DNet network that integrates the idea of dense connections to obtain multi-scale features of local image data within a prior frame, improving the accuracy of fault classification; Finally, Bayesian algorithm is used to improve the model hyperparameters to obtain the optimal combination of learning rate, convolution kernel number, etc., to achieve efficient and accurate fault identification and classification. The research results show that： Compared with the original algorithm, the improved YOLOv4-Tiny algorithm improves the accuracy rate by about 5.3%, and the improved Res_DNet algorithm improves the accuracy rate by more than 4.6% compared with the classical algorithm, which can realize the high-precision identification of thermal defect status of substation equipment