Abstract: Aiming at the problem of missing AIS trajectory number of inland waterway trunk ships, a method of ship trajectory data repair method based on bidirectional long-term and short-term memory network (BLSTM-RNN) model was proposed. A two-layer bidirectional recurrent neural network (RNN) model was constructed by using ship trajectory context information and other return features as model input. In the model input, correlation analysis and sequence autocorrelation coefficient were used to determine the correlation variables of ship trajectory points and the autocorrelation lag value of ship trajectory sequence. In the model structure, the super-parametric values of the model were reasonably set with the ACC rate as the index. Taking the ship trajectory data of Wuhan section and Chongqing section of the Yangtze River trunk waterway as samples, empirical verification of the proposed model was carried out. The experimental results show that, compared with linear and other machine learning methods, the accuracy of BLSTM-RNN method is improved to a certain extent. In the experiment of straight-line reach in Wuhan section, the repair error is controlled within 15 m magnitude, which is much lower than that of other non-linear methods. In the complex reach of Chongqing section, the repair error can be controlled in the order of 10m. In addition, the proposed model solves the problem of accuracy loss of traditional methods at long-distance loss points, and reduces the repair error to 50 m magnitude when 20 continuous points are lost.

Key words: marine engineering, bidirectional long-term and short-term memory network (BLSTM), recurrent neural network (RNN), ship trajectory restoration, ship autopilot

摘要： 针对内河干线船舶AIS轨迹数缺失问题，提出一种基于双向长短时记忆网络(BLSTM-RNN)模型的船舶轨迹数据修复方法。通过利用船舶轨迹上下文信息及其他回传特征作为模型输入，构建两层的双向循环神经网络(RNN)模型。在模型输入上，采用相关性分析及序列自相关系数，确定船舶轨迹点相关变量及轨迹序列自相关滞后值；在模型结构上，以ACC率为指标对模型超参数值进行合理设置，以长江干线航道武汉段及重庆段船舶轨迹数据为样本，对模型进行实证验证。实验结果表明：与线性及其他机器学习方法相比BLSTM-RNN方法在精度上有一定提升；在武汉段顺直河段实验中，将修复误差控制在15 m量级内，远低于其他非线性方法的50 m量级；在重庆复杂河段内，可将修复误差控制在10 m量级；模型解决了传统方法在长距离丢失点上精度缺失的问题，在20个连续点丢失的情况上，将修复误差降低至50m量级。

关键词: 船舶工程, 双向长短时记忆网络(BLSTM), 循环神经网络(RNN), 船舶轨迹修复, 船舶自动驾驶