The coverage of the communication network in the railway tunnel mainly depends on the communication leaky coaxial cable (leaky cable). The leaky cable is usually suspended on the tunnel wall and fixed by the leaky cable fixture. Affected by the tunnel environment and train operation, fixture failures and shedding occur from time to time, which seriously threatens the safety of operation. With the advent of the era of artificial intelligence, the detection technology based on deep learning provides a new idea for the state detection of leakage cable fixtures in tunnels. This project relies on the efficient YOLOv5 framework to improve, add Ghost convolution and deep separable convolution to reduce the amount of calculation of the model, and add SE attention mechanism to avoid the decline of accuracy, innovative proposed a GS-YOLOv5 tunnel card detection algorithm. Compared with other models, the speed changes little, but the accuracy is the best, especially for damaged fixtures, the accuracy is as high as 98.4%, which has obvious advantages. On this basis, this project designs and implements a tunnel fixture detection system based on deep learning. The system includes client identification module, hardware detection equipment, data visualization platform and management module. The system has a complete structure and reasonable blocks. It is automated from data collection to detection and identification results, accurate and efficient, and meets the needs of real-time detection. It can provide theoretical reference and technical support for railway informatization and intelligence. The main workflow of this project is to first draw the data collected by CMOS high-speed camera into space-time diagram, then binarize the space-time diagram and construct convolution filter to denoise, then use vertical projection combined with double mark bit method to correct errors to realize accurate extraction of tunnel boundary, finally use GS-YOLOv5 algorithm to detect, classify and mark faults of fixtures, and timely feedback the data to the back end. The hardware part deploys the detection model to the hardware, performs real-time detection on site, and feeds the results back to the back-end, which processes the data, then presents the data to the staff and assigns personnel and tasks through the management system. (1) Client interface construction: In order to facilitate user identification, PYQT5 is used to build a client identification interface, which has operations such as picture carousel, result display and identification, and operations such as carousel, identification and progress bar update. At the same time, consider the customer's use experience, add multi-thread processing, avoid the situation that the program is stuck due to the user's operation error in the identification process, and have a progress display page to remind the user to detect the progress, which is more clear and clear. The inspection personnel can check the inspection image of a specific road section according to the requirements, and judge the abnormal fixture status twice. (2) Construction of data visualization platform: In order to better present the detection results to managers, the project designs a data visualization platform, including personnel display, route display, fixture display and other information. Through Echarts rendering, the platform can simultaneously retrieve statistical information of different types of fixture status on the same road section, and display the damage status and change trend of fixtures. The back-end writes the interface according to the required information, and the front-end writes the UI interface according to the back-end interface. (3) Construction of management platform: in order to carry out the subsequent fixture maintenance work in an orderly manner and show the data to users, the project has built a management platform to visualize personnel information, line information, fixture information, department information, task information, etc. and allow staff to carry out corresponding processing operations. The back-end writes corresponding interfaces. The front-end requests the back-end interfaces to retrieve data and display them through VUE rendering, comprehensive inspection and centralized management of leaky cable fixtures are realized. (4) hardware deployment: due to the problem of clip picture segmentation in software detection, resulting in missed detection and false detection, this project is supplemented by hardware system design. the hardware uses jetson nano bo1 motherboard with CMOS industrial camera, which can be fixed on the detection vehicle and other carriers to shoot and obtain images. A display may be used for the demonstration of the detection process and results. The hardware adds the Deepsort tracking algorithm on the basis of YOLOv5 model to avoid the same fixture being repeatedly identified and reduce data redundancy. Because the detection speed will slow down after adding the tracking algorithm, the model is accelerated by TensorRT, and the detection rate reaches 45.5 frames per second, which meets the needs of real-time detection. In view of the application scenario of tunnel leakage cable fixture fault detection and identification, the detection method is improved by combining software and hardware. In terms of software testing, a fixture testing platform, a fixture information visualization platform and a fixture maintenance personnel management platform have been set up. The data of each platform can be shared and updated synchronously in time. In order to avoid the problem that the same fixture is divided in the process of software image detection, the hardware detection system is developed, combined with the Deepsort tracking algorithm, and the TensorRT technology is used to speed up the detection speed, which can be used for on-site real-time detection. The whole set of detection system includes client identification module, hardware detection equipment, data visualization platform and management module, the system structure is complete, the block is reasonable, from data acquisition to detection and identification results automation, accurate and efficient equipment cost is low, has broad application prospects.