为准确甄别船舶身份，提出一种基于层次注意力孪生网络的船舶身份甄别模型，结合时间注意的长短期记忆和多尺度的卷积网络，从时间和语义信息层面对船舶轨迹进行表征，并采用以改进的孪生神经网络计算船舶轨迹表征向量间的差异度作为判断船舶身份的依据。为验证该模型的有效性，在厦门港及附近水域船舶轨迹数据基础上对比分析本文模型和常用机器学习模型甄别船舶身份的性能。结果表明，本文模型可在小规模数据集上学习得到泛化性较好的船舶身份甄别性能，在测试集上的 F₁分数为0.8971，而常用机器学习模型在相同测试集上仅能达到0.7774，可见，本文模型可满足船舶身份甄别和异常排查等相关应用的需要。

In order to accurately identify the ship’s identity, a ship identity recognition model based on hierarchical attention Siamese network was proposed. By combining the long-term and short-term memory of temporal attention and the multi-scale convolutional network, the ship trajectory has been characterized at the level of temporal and semantic information, and by using  an improved Siamese network to calculate the dissimilarity score between the ship trajectory characterization vectors as a basis for determining the ship identity. To verify the effectiveness of the proposed model, the performance of the proposed model and the commonly used machine learning and deep learning models were compared and analyzed in terms of the ship trajectory data of Xiamen port and nearby waters. The results show that the proposed model can obtain good generalization performance of ship identity recognition on a small-scale data set, with a F₁ score of 0.8971 on the test set, while the commonly used machine learning models can only achieve a F₁  score of 0.7774 on the same test set, which indicates that the proposed model can meet the needs of applications related to ship identification and anomaly detection.