Abstract: Underwater acoustic communication signal recognition is an important part of underwater acoustic communication reconnaissance and an important prerequisite for implementing guided communication countermeasures. Traditional underwater acoustic communication signal type identification methods are usually based on pattern recognition technology and rely on the professional knowledge and experience of domain experts for feature selection and extraction, which are highly subjective; intelligent identification methods based on deep learning can achieve feature extraction and identification. Integration, but currently network training is mainly based on a single spectrum. When the signal has little difference in a single feature, all signal types cannot be correctly classified based on a single spectrum. In order to solve this problem, an intelligent method for underwater acoustic communication signal type identification based on multi-spectral graph joint of VGG16 (Visual Geometry Group 16)network is proposed. By performing time-frequency analysis and quadratic spectrum analysis on the signal, the joint spectrum of the time-frequency spectrum and the quadratic spectrum is used as network input to construct a signal data set for training. After the network training loss and accuracy meet expectations, Testing of signal type identification can be implemented. Six types of signals, including Linear Frequency Modulation (LFM), Binary Phase Shift Keying (BPSK), Quadrature Phase Shift Keying (QPSK), 2-Frequency Shift Keying (2FSK), 4-Frequency Shift Keying (4FSK), and Orthogonal Frequency Division Multiplexing (OFDM), were processed using simulation data and experimental data collected on a lake. The results show that the intelligent recognition method based on a single time-frequency diagram cannot distinguish between BPSK and QPSK signals, leading to a decrease in the overall recognition rate. In contrast, the intelligent identification method based on the combination of multi-spectral images can make correct classifications, and the accuracy of identifying various types of signals exceeds 99%, thus verifying the feasibility of the proposed algorithm.