Abstract: Time-frequency domain reflectometry relies on human experience judgment when locating and diagnosing the damp level of cable joint. Therefore, this paper proposes a method of diagnosing damp defect of cable joint based on prior knowledge driven neural network, which realizes the intelligent prediction of cable joint damp level through neural networks. The time delay, peak voltage and phase characteristics of the time-domain signal at the defect are extracted as the input of the diagnosis network. The correspondence between the signal characteristics and the degree of dampness of the cable intermediate joint is established through the fully connected network. The model is trained by simulation data and verified by simulation data and experimental data. The results show that the diagnostic network constructed in this paper can effectively predict the degree of joint moisture.