基于神经网络和AIS数据的海上船舶分类识别方法研究

The Automatic Identification System (AIS) is an information exchange platform for vessel-to-vessel and vessel-to-shore or satellite communications for marine vessels.According to the vessel trajectory data provided by AIS, the recognition of vessel targets can be used not only for monitoring and management of maritime affairs in civilian place such as commercial ports, but also for discovery, tracking and early warning for military purposes.In the project, the vessel trajectory data provided by AIS is used as the data source to study the noise processing and feature extraction of the vessel trajectory data, and recognize different vessel types through the extracted trajectory features. The research contents include: (1) Studying the noise algorithm which is applicable to AIS data, providing the high-quality data for vessel recognition, reducing the impact of noise on vessel classification accuracy; (2) Establishing deep auto-encoder model based on the characteristics of trajectory data, introducing the attention mechanism in machine learning to mine the motion features implied in the trajectory data; (3) Adopting the integrated learning model named XGBoost and taking the trajectory features extracted by auto-encoder as input to recognizing the vessels, such as the cargo vessels, fishing vessels, yachts and warships. Through the research of this project, the trajectory data mining theory and method can be deepened and expanded, and it is of great significance for promoting the practical application of trajectory data.

船舶自动识别系统（AIS）是为海上船舶提供的一种船舶间以及船舶和岸基通讯或卫星通讯的信息交互平台。根据AIS提供的船舶轨迹数据，对船舶目标进行分类识别，不仅可以用于民用场所的海事监控和管理，还可以用于海面军事目标的发现、跟踪和预警。本项目以AIS产生的船舶轨迹为数据源，研究船舶轨迹数据的噪声处理和特征提取，并通过提取的轨迹特征识别不同的船舶类型。研究内容包括：（1）研究适用于船舶轨迹数据的噪声算法，为船舶分类识别提供高质量的数据，降低数据噪声对船舶分类精度的影响；（2）根据轨迹数据的特点，建立深度自编码器模型，引入机器学习中的注意力机制，挖掘轨迹数据隐含的运动特征；（3）采用XGBoost集成学习模型，以自编码器提取的轨迹特征为样本，对货船、渔船、游艇和军舰等船舶进行分类识别。通过本项目研究，能够深化和拓展轨迹数据的挖掘理论和方法，对于推动船舶轨迹数据的实际应用具有重要意义。

基于AIS数据对船舶进行分类识别在保障海上交通安全中发挥着重要作用。但是，船舶轨迹在宽阔的海上复杂多变，船舶特征提取和分类建模存在一定难度，因此迫切需要一套完善的技术体系来解决当前基于AIS数据船舶分类识别研究中的难题。本项目围绕“海上船舶分类识别”这个主题，以AIS数据作为研究对象，采用机器学习方法，展开如下3个方面的研究：（1）在已有轨迹噪声处理研究方法基础上，提出一种新的基于规则模型和密度聚类的轨迹虑选方法。（2）参考近年来轨迹特征提取方法，提取轨迹段中的速度、加速度、航向、航向变化率以及离岸距离等五大类来构建特征空间。进一步，将提取的轨迹特征作为输入量，提出一种基于集成学习的船舶目标分类识别方法。（3）针对利用轨迹数据对海上船舶目标分类的传统方法多数依赖于人工提取特征，且难以兼顾空间和时间特征对分类结果的影响等问题，提出一种卷积神经网络与双向门控循环单元相结合的组合分类模型。本项目提出了一套系统的基于机器学习和AIS数据的船舶分类识别方案，包含从原始AIS数据预处理到船舶运动特征提取、到模型构建、再到船舶分类识别的完整技术流程。依托本项目，课题组共发表学术论文8篇，已录用学术论文2篇，其中EI检索3篇，申请国家发明专利3项。