1.National Space Science Center;2.College of Systems Engineering
航迹预测是保障船舶航行安全、提高海洋交通管制效能、高效搜索海面目标的关键技术。为提高船舶航迹预测精确度，针对航迹特征多维度的特点，本文提出了一种并行LSTM-FCN（Parallel LSTM-FCN，PLSTM-FCN）模型。该模型有效结合了LSTM模型对时间序列数据长期趋势预测的优势和FCN模型擅于提取时间序列数据细节变化规律的特点，通过并行结构设计保证相同训练效率下提取特征参数翻倍，实现了较高精确度的高维航迹数据特征提取和趋势预测。提出了基于动态时间规整算法和拉依达准则的船舶历史航迹数据预处理方法，提高了PLSTM-FCN模型从不同类型船舶历史航迹中深度学习航行趋势和转弯细节的效率。开展了基于船舶自动识别系统（Automatic Identification System，AIS）数据的仿真实验，实验结果表明PLSTM-FCN模型对多维特征船舶航迹预测的精确度明显优于传统循环神经网络。
Trajectory prediction is very important to navigation safety, marine traffic control and surface vessels search. In order to improve the accuracy of vessel trajectory prediciton and according to the multi-dimensional characteristics of vessel trajectory features, a new model named Parallel LSTM-FCN(PLSTM-FCN)is proposed. The model can exact features and trend from multi-dimensional vessel trajectory, because of combining with the LSTM which has advanced to predict time series trend and the FCN which is adept in exacting detail features of time series. Simultaneously, the training efficiency of PLSTM-FCN which has more parameters is the same as LSTM-FCN, because of the concurrent design. In order to improve the learning efficiency, a preprocessing method based on dynamic time warping algorithm and Laida criterion is proposed. The simulation experiment is carried out based on the data of Automatic Identification System (AIS). Experimental results show that the PLSTM-FCN is more accurate than typical RNN in vessel trajectory prediction.