船舶航行交通事件实时检测技术研究现状与展望

船舶航行交通事件检测依赖基于历史数据的离线检测方法, 检测模型适用性差, 难以满足监管人员的实时监测需求。通过分析船舶异常行为检测、航行事故检测等现有交通事件检测技术, 可以发现: 在数据层面, 监测数据来源单一、环境信息缺失; 在方法层面, 基于统计、风险评估等经典模型的事件监测方法效率高但准确性低, 基于神经网络、图像识别等机器学习的检测方法准确性高但效率低; 多源数据融合、多项技术结合的交通事件检测方法成为实时检测方法的发展趋势。在此基础上, 梳理了实时船舶航行交通事件检测的3项关键技术: (1)海事大数据技术: 高效处理船舶运动数据和航行环境数据, 统一多源异构数据结构标准, 降低数据源单一造成的事件误报率; (2)船舶行为动态建模技术: 利用知识图谱等技术融合船舶航行情境信息, 在不同船舶运动环境下利用深度学习、语义关联、图神经网络等方法构建不同的船舶行为模型, 提高检测准确性; (3)实时分析和可视化技术: 结合平行系统进行虚实系统间信息传递, 定性分析检测结果, 实时显示检测全过程, 提升监管过程中的人机交互效率。然后, 提出了包括数据采集、后台服务和客户端应用3个功能模块的交通事件平行检测系统; 该系统具备实时接收并处理船舶航行数据、分析并预测交通状态、动态检测并预警交通事件和仿真结果展示等功能。从数据融合、交通状态感知和交通虚实映射3个方面, 展望了面向海事监测实务的实时检测技术发展方向。 

A Real-time Detection of Nautical Traffic Events: A Review and Prospect

Nautical Traffic Event Detection(NTED) methods mostly rely on offline methods using historical data, which are insufficient for real-time traffic supervision. The studies on abnormal behavior detection and incidents detection of ships are collected and investigated, and the findings are concluded as follows: from the perspective of data, the detection data rely on a single source and the environmental information is usually missing; from the perspective of methodologies, classical models that are based on statistical methods, risk assessments, etc., have high efficiency but low accuracy; while, the machine-learning based methods, such as neural networks, image recognition, etc., have high accuracy but low efficiency; and the combination of multi-source data fusion and multi-technology have become new trends.Three key technologies for the real-time NTED are summarized: ① maritime big data technologies, which process ships and environment data efficiently and standardize multi-source heterogeneous data structures, which reduces the false alarm caused by the single data source; ②dynamic behavior modeling, which uses knowledge graph or other technologies to integrate nautical contextual information, and uses deep learning, semantic association, graph neural network or other methods to develop different models for dynamic ship behaviors in different nautical context, which improves the accuracy of the NTED; ③the real-time analysis and visualization techniques combined with parallel systems, which can transfer information between the virtual and real systems, analyze the simulated results, and display the detection process which facilitates human-computer interactions in the supervision. A Nautical Traffic Event Parallel Detection System(NTEPDS) is proposed, which includes three functional modules: ①the data acquisition; ②the backend service; ③the client application. The NTEPDS can receive real-time navigation data, analyze and predict real-time traffic status, dynamically detect and report traffic events and display the simulation results. Finally, the prospects of the real-time NTED are concluded from three aspects: data fusion, traffic state perception, and traffic virtuality-reality mapping, which reveals the development directions of real-time NTED at the practical level.