There are still some problems in the detection of insulators in substations, such as low accuracy, relying on artificial proximity to insulators for detection. The team designed the substation insulator intelligent inspection system, the project is mainly composed of autonomous tracking part, inspection machine part, PC-side algorithm model recognition part. Automatic tracking is carried out in the substation through the intelligent inspection machine, the industrial camera carries out visual identification of the insulator, the collected image is transmitted to the PC side through the digital image, and the image processing is carried out with the TensorFlow deep learning algorithm to accurately identify the problems existing in the insulator, so as to achieve the purpose of automatic detection of the insulator in the substation. The intelligent inspection machine takes the STM32 controller as the control core, and is composed of the body movement part and the camera identification transmission part. The car body movement part includes 4 12V motors, 4 mecanum wheels, 4 encoders for measuring the real-time speed of the motors, a 32 single-chip main control board, an OLED screen for displaying the current task and position, a BTN motor drive module and a 12V9AH battery. The camera identification transmission part includes a KS2A42 industrial camera identification module, DJI-FPV image transmission module and two MG996 steering gear. The camera is installed on the gimbal mounted by two MG996 steering gears. The two steering gears can collect and feedback information through the MPU6050 to further adjust the best shooting angle. The camera is connected with the DJI-FPV image transmission module. The inspection machine automatically takes photos after arriving at the designated area, collects the insulator image and sends instructions to the DJI-FPV, and the DJI-FPV receives the instructions and transmits them to the PC. When training the network model, keras, a deep learning library under the TensorFlow framework, is used to extract vgg16 intermediate results for transfer learning, then loss is redefined as the parameters of the neural network and the loss value of cross entropy in the model. During the training process, the number of network parameters is optimized, with a total of 36000 training times. The training accuracy rate reaches 0.9815 and the test accuracy rate reaches 0.93788. It solves the shortcomings of high memory occupancy rate and slow detection speed caused by the huge parameters of vgg16. In the aspect of autonomous tracking, the automatic tracking system of UWB RFID is used. The intelligent inspection machine locates through UWB base station, and navigates through RFID and magnetic track radio frequency identification. It adapts to the complex environment of strong electromagnetic interference in substation, and has the advantages of high positioning accuracy and strong anti-interference ability.