Abstract: Objectives To address the issues of low recognition accuracy and weak robustness in traditional underwater acoustic target recognition methods within complex marine environments, this paper proposes a method based on a Convolutional Window Attention Network. Methods First, Mel-frequency cepstral coefficients were used to preprocess underwater acoustic signals to obtain an MFCC feature matrix. Then, a convolutional embedding module was designed to extract multi-scale features from the MFCC feature matrix through multi-layer one-dimensional convolution operations with progressively expanded receptive fields. Next, a self-attention module was constructed to model the global correlations of multi-scale features by partitioning local windows and assigning attention weights. In addition, a shifted window attention module was introduced to further capture long-range dependencies among features by shifting and reorganizing the feature sequence. Finally, underwater acoustic target recognition was achieved through a linear classifier. Results Validation on the DeepShip dataset demonstrates that the constructed model achieves an accuracy of 97.68%, outperforming other comparative models. Conclusions The proposed method demonstrates good robustness and can effectively achieve underwater acoustic target recognition under low signal-to-noise ratio conditions.