Abstract: To address the problems of complex fault types in coaxial cable operation, insufficient localization accuracy of traditional time-frequency-domain detection methods, and low efficiency of manual identification, a deep learning-based coaxial cable fault identification and localization system was developed. The overall detection scheme and algorithm framework of the system were designed in accordance with relevant industry standards and engineering application requirements for coaxial cable fault detection. A high-sensitivity signal acquisition module was developed to ensure fine frequency-domain feature extraction. Frequency-domain signal preprocessing, a lightweight deep learning network, and fault localization algorithm were effectively integrated, so that the system was enabled not only to measure fault location but also to automatically determine fault type, whereby efficiency and accuracy of fault classification were improved. An experimental test platform was constructed to evaluate the performance of the system. It was verified by experiments that various fault types, including low-resistance faults, high-resistance faults, broken-line (open-circuit) faults and connector faults, could be effectively identified; a relative localization error of less than 0.14% and a comprehensive fault identification accuracy of 98% were achieved, and localization error was shown to meet engineering requirements. The system was demonstrated to be suitable for intelligent operation, maintenance and fault diagnosis of coaxial cables in communication, radio-frequency and other scenarios.