Modulation Recognition of Underwater Acoustic Communication Signals Based on Dual Spectral Fusion

A dual-spectrogram feature fusion-based modulation recognition method is proposed for blind modulation recognition of underwater acoustic communication signals under prior knowledge deficiency. It is discovered that passband single-carrier (SC) and orthogonal frequency division multiplexing (OFDM) signals exhibit statistical characteristics approaching Gaussian distribution after shaping filtering, which renders conventional higher-order spectral analysis ineffective due to its low sensitivity to Gaussian-like signals. To address this, a joint time-frequency and higher-order spectral analysis framework is established: Time-frequency diagrams are extracted through continuous wavelet transform to achieve coarse classification by leveraging the disparity between the time-frequency aggregation of SC signals and the spectral expansion of OFDM signals. To compensate for the inadequate phase sensitivity of time-frequency diagrams in distinguishing BPSK/QPSK signals, a dual-spectrogram fusion framework is constructed by integrating higher-order spectral diagrams that capture phase nonlinearity characteristics. A lightweight neural network model embedded with the Convolutional Block Attention Module (CBAM) is designed to optimize feature extraction channels, enhancing the model's focus on dual-spectrum phase coupling features and marginal band characteristics in time-frequency diagrams under low signal-to-noise ratios (SNRs). Simulation results demonstrate that the proposed method achieves 91% recognition accuracy for BPSK, QPSK, 2FSK, 4FSK, and OFDM signals at 4 dB SNR, outperforming traditional single-spectrogram approaches by approximately 10%.