Abstract: Traditional monitoring methods have obvious deficiencies in real-time performance, accuracy, and early warning capability, making it difficult to effectively cope with the complexity of the underground cable operating environment. To address the challenges of early fault prediction for underground cables, this paper proposes a physics-enhanced CNN-LSTM deep learning architecture, which realizes real-time monitoring and early warning of early faults by analyzing cable vibration signals. First, the physical characteristics of cable vibration signals are analyzed in depth, and the prior knowledge of cable mechanical dynamics is embedded into the deep learning model to construct a hybrid model integrating the feature extraction capability of convolutional neural networks (CNN) and the temporal modeling capability of long short-term memory (LSTM) networks. An underground cable simulation platform is built, and comparative experimental analysis is conducted through different models such as SVM, Random Forest, CNN, LSTM, CNN-LSTM, CNN-LSTM+Attention, PINN, PhyCNN, and physics-enhanced CNN-LSTM. The results show that the average recognition accuracy of the physics-enhanced CNN-LSTM model for six types of typical cable faults reaches 96.8%, which is 7.2% higher than that of the traditional CNN-LSTM model, and the false alarm rate is reduced to 2.1%. In addition, by introducing a physics-constrained loss function, the generalization ability of the model in scenarios with a small number of samples is significantly improved, providing an effective technical approach for the predictive maintenance of underground cables.