A leakage pressure identification model in the water supply pipeline was proposed to resolve problems that the existing studies of leakage detection failed to determine the pressure of leakage. The proposed model was based on the short-time Fourier transform with window parameter optimization and convolutional neural network. Pipeline sound signals were collected from a full-scale pipeline leakage pilot experimental platform. Leakage conditions of the same leak area but different shapes were subjected to short-time Fourier transform processing to obtain two-dimensional time-frequency spectra. The spectra contained characteristics related to leakage at three different pressure levels and were used as input for the convolutional neural network. The leakage pressure identification model was constructed based on the optimization of window parameters and network hyperparameters. Experimental results showed that the proposed model achieved an overall accuracy of 95.2%, and the recognition of high-pressure, middle-pressure and low-pressure conditions were 93.5%, 92.9% and 92.4% respectively. The proposed model has a higher accuracy in identifying leakage and pressure of leakage than traditional machine learning models, and the model can be used for identifying leakage pressure levels in real water supply networks.