Abstract: Underwater acoustic signal classification is an important research direction in underwater acoustics, and an effective feature extraction and classification decision method is of great concern for underwater acoustic signal classification technology. In this paper, based on wavelet packet decomposition, the time-frequency map features of three kinds of underwater acoustic signals, namely fish vocalization signals, merchant ship radiated noise signals and wind-generated noise signals, are extracted, and a convolutional neural network (CNN) with seven-layer structure is set up as a classifier. The research results show that the accuracy of the three kinds of acoustic signals by the methods of combining the time-frequency map features of wavelet packet and the convolutional neural network can reach (98±1)% in different test sets. Therefore, the method can be expected to be used for the classification of more underwater acoustic signals. This research results can provide a reference for the classification and recognition of underwater acoustic signals.