A Novel Noise Learning Model for Underwater Acoustic Signal Denoising Based on Local–Global Convolution-Enhanced Transformer

A novel underwater noise learning model with a local–global convolution-enhanced transformer (NL-LGCET) is proposed to underwater acoustic signal denoising within complex marine environments. The model introduces a local–global convolution-enhanced transformer (LGCET) with a dual-branch mechanism for parallel extraction of local and global temporal features, surpassing traditional serial structures. A convolution-enhanced transformer (CET) block integrates multihead self-attention (MHSA) with an advanced depthwise separable convolution to capture intricate time–frequency features. Replacing the traditional short-time Fourier transform (STFT), a learnable 1-D convolution module preserves phase characteristics and adapts to environmental noise for precise feature extraction. Additionally, a noise learning module and time-domain residual mechanism ensure accurate noise estimation and signal reconstruction. Experimental results shows that the proposed model achieves 19.11 dB scale-invariant signal-to-noise ratio (SISNR) and 20.40 dB signal-to-distortion ratio (SDR) on the ShipsEar dataset, representing significant improvements over the optimal baseline. Furthermore, the model exhibits robust generalization capabilities in unknown ship scenarios and real scenarios.