Cable faults threaten the safe and stable operation of smart grids, and vibration signal diagnosis research on cables based on artificial intelligence technology can effectively enhance the reliability of smart grids. In order to improve the speed and accuracy of cable defect identification, this paper proposes a partial discharge identification method for cables based on a fully convolutional bidirectional long short-term memory neural network (BiLSTM-FCN). The time-domain characteristics of different working conditions under industrial frequency AC voltage are collected by building a vibration signal-based partial discharge test platform for cables. The Overlap strategy enhances the data set. The BiLSTM layer is introduced to process the one-dimensional time domain waveform signal, and combined with the local detail features extracted by the FCN layer, the data features are more abundant. Thus, a more accurate diagnosis of localized discharge in high voltage cables under different operating conditions. The results show that the diagnostic accuracy based on the BiLSTM-FCN model reaches 92.2%, which is 1.2% and 1.4% higher compared to the FCN model and the BiLSTM model. BiLSTM-FCN model possesses a better recognition effect and faster recognition speed in the identification of partial discharge defects type, which can effectively achieve the automatic detection of abnormal fault nodes of smart grid cables. It is significant for realizing online dynamic evaluation of small portable online monitoring equipment and provides a reference for future related research.