A cable fault detection method and device provided by the embodiments of the invention belong to the field of electric power. The method includes the following steps: carrying out feature extraction on the current signal of a cable to be processed to obtain primary features; identifying the primary features according to a preset deep learning network model to get effective features; and classifying and identifying the effective features to get a classification result used to characterize whether the cable to be processed has a fault. According to the method, after the primary features are extracted, the primary features are identified to obtain effective features, and the current signal of the cable to be processed is classified and identified. Therefore, the identification accuracy of early fault of cables is improved, and accurate detection and identification is realized.