Power cables are usually directly buried or installed in the cable channel for laying, and the intermediate connector is the weak link in the cable, which is more susceptible to water intrusion and failure, seriously affecting the operation safety of the transmission and distribution system. Therefore, a deep learning network based on time-frequency domain reflection method (TFDR) for automatic detection of cable joint moisture is proposed in this paper. The network structure consists of two parts. In the first part, variational autoencoder (VAE) network is used to detect the time domain signal of the time-frequency domain reflection method, so as to identify and locate the damp cable joint. In the second part, convolutional neural network is used to extract the signal features of the damp junction, and then the degree of damp is judged by the fully connected neural network. Finally, a 40 m 10 kV XLPE single-phase power cable test platform was built for verification, and the automatic detection of cable joint moisture was successfully realized.