A Two-Stage Magnitude-Complex Spectrum Attention Model for Self-Noise Suppression in Shipborne Sonar Platforms

Self-noise suppression technology for shipborne sonar platforms is designed to mitigate the inherent challenge of significant self-noise interference during sonar detection. This is achieved by suppressing the self-noise generated by the sonar platform itself. Traditional underwater signal suppression algorithms often fall short in effectively suppressing platform self-noise, particularly in scenarios with low signal-to-noise ratios. To address this problem, we propose a two-stage magnitude-complex spectrum attention network model (MP-CPNet) for self-noise suppression in shipborne sonar platforms. The model comprises a magnitude processing network (MPNet) and a complex processing network (CPNet). The former part inputs the signal's magnitude and performs a coarse recovery, then coupling the output magnitude with the phase of the noise signal to obtain a coarse estimated signal. The latter part accepts the coarse estimated signal and noise spectrum as input to further estimate the spectral details, then performing a fine recovery of the signal phase and magnitude. The result estimated by the entire network is then output. Each sub-network includes multiple improved self-attention modules, which can learn and model the long-term dependencies of features in time and frequency, thereby promoting overall spectrum recovery. Experimental comparisons demonstrate that MP-CPNet outperforms traditional underwater noise suppression algorithms and other deep learning-based algorithms in terms of suppression effects.