摘要 : Ship-radiated noise recognition is an essential but complex task in the construction of marine information systems and marine scientific research. Ambient noise, unstable frequency shifts, and irregular multipath interference make... 展开 Ship-radiated noise recognition is an essential but complex task in the construction of marine information systems and marine scientific research. Ambient noise, unstable frequency shifts, and irregular multipath interference make it complicated to recognize ship-radiated noise accurately. Existing recognition methods exhibit constrained proficiency in the identification of the motion states of ships, thus leading to disappointing application performance. To effectively recognize the ship movement with less computation, this work proposes the time-frequency Swin-Transformer (TFST) network. A hierarchical self-attention module is presented to extract multilayer time-frequency features so that the TFST network could learn moving targets' features in TF representations of the noise radiated by moving targets. A scale-difference simplified architecture is designed to reduce network complexity. Experiments reveal that the TFST network outperforms the state-of-the-art convolutional neural networks (CNNs) and Transformers on two underwater acoustic data sets. Moreover, the TFST network achieves at least 1.3 times improvement compared to five state-of-the-art methods on both average accuracy (OA) and kappa coefficient in three motion status recognition experiments. 收起