Objective Complex marine environments are characterized by various types of mixed interference, including ocean current noise, marine biological noise, and ship-radiated noise. Traditional underwater acoustic target recognition methods rely heavily on manually engineered features and lack the capacity to effectively extract multi-scale temporal features, resulting in limited classification accuracy and poor robustness under low signal-to-noise ratio (SNR) conditions. To address these limitations, this paper proposes a Convolutional Window Attention Network (CWA-Net) for underwater acoustic ship target recognition.
Method Mel-frequency cepstral coefficients (MFCC) are employed as preprocessing features to extract spectral envelope information from raw underwater acoustic signals. A multi-layer one-dimensional convolutional embedding module is developed to progressively enlarge the receptive field and extract multi-scale local temporal features. A window-based multi-head self-attention (W-MSA) block is constructed to model global feature correlations within local non-overlapping windows. Furthermore, a cyclically shifted window multi-head self-attention (SW-MSA) module is introduced to facilitate cross-window information exchange and capture long-range dependencies among feature sequences. Finally, a linear classifier is employed to predict target categories. Experiments are conducted on the DeepShip dataset, which contains underwater acoustic data from four vessel categories. Gaussian white noise with different SNR levels is added to simulate noisy marine environments. Representative baseline models, including ResNet, TCN and Swin Transformer are set for performance comparison. In addition, ablation studies and parameter sensitivity analyses are performed to investigate the contribution of each component.
Results The proposed CWA-Net achieves an overall recognition accuracy of 97.68% on the original dataset with only 2.874 M parameters. Its classification accuracy remains above 90% under SNR conditions ranging from 5 dB to −10 dB. T-distributed Stochastic Neighbor Embedding (t-SNE) visualization demonstrates that the feature embeddings generated by CWA-Net exhibit compact intra-class clusters and clear inter-class boundaries. Ablation studies confirm that the preprocessing module, convolutional embedding layer and dual attention blocks all contribute significantly to performance improvement, and optimal configurations exist for convolution kernel size, attention window size and number of stacked attention layers.
Conclusion CWA-Net simultaneously realizes multi-scale local feature extraction and global long-range dependency modeling. It owns lightweight model scale and strong anti-noise robustness, and can effectively recognize ship underwater acoustic targets under low-SNR marine environments, which provides practical value for marine monitoring engineering applications.
DownLoad: