Sonar pulse recognition refers to the process of extracting features from underwater sonar pulse signals and classifying targets. The main process includes feature extraction and target classification. However, the marine environment is complex, with reverberation, multipath effects, and interference noise from other ships or marine organisms, which greatly reduces the effectiveness of traditional sonar pulse detection and recognition. Deep learning takes the artificial intelligence neural network as the backbone, and it consists of multiple processing layers to study data with different orders of magnitude. A deep learning network model can process structured and unstructured data, while avoiding automatic feature extraction by manual operation. When the amount of data reaches a certain level, deep learning methods can effectively achieve precise recognition of sonar pulses. However, due to the difficulty of underwater data collection and high experimental costs, the size of the sonar pulse recognition training set is a key issue affecting recognition. Starting from introducing the structure of deep neural networks, this paper proposes a method for generating sonar pulse recognition training sets based on DCGAN (Deep Convolutional Generative Adversarial Networks).