交汇水域船舶轨迹预测与航行意图识别

针对典型水上交通场景交汇水域，研究了1种数据驱动的船舶轨迹预测与航行意图识别方法。设计CNN+LSTM组合神经网络，通过学习交汇水域船舶的历史轨迹，以CNN+LSTM网络为编码器提取其通航环境及船舶航行时空特征，LSTM与全连接层为解码器同步输出未来时段内船舶轨迹序列和航路选择，从而形成船舶轨迹与航行意图识别模型。同时，引入Dropout网络结构描述该模型的预测不确定性，采用随机关闭CNN+ LSTM核心网络部分神经单元的方式，以相同轨迹序列作为输入获取多组相近的预测结果，根据其统计均值与方差对船舶轨迹预测的不确定性进行量化。以美国沿海某交汇水域公开AIS数据为对象开展实验，创建了该交汇水域船舶航行轨迹数据集，以输入时长60 min，采样频率3 min作为输入条件，Dropout值取0.5，实验结果表明：所提方法对未来60 min时段内的轨迹预测误差为3.946 n mile，航行意图识别准确率达87%，不确定性估计覆盖率达85.7%。与LSTM预测方法相比，当船舶操纵性发生改变时，所提CNN+LSTM模型的轨迹预测误差降低了31.6%，而且兼具船舶航行意图识别及预测不确定性估计能力，有利于智能航行与海事监管技术发展。

Trajectory Prediction and Intention Identification of Ships in Confluence Waters

A data-driven method is proposed for predicting ship trajectory and identifying sailing intention for typical confluence waters. A CNN+LSTM combined neural network is designed by learning the historical trajectories of the ships in the confluence waters. Using CNN+LSTM as an encoder, the spatio-temporal characteristics of the navigable environment and ship sailing are extracted. The decoder, which is composed of the LSTM and full connection layer, synchronously outputs the trajectory sequence and route selection of the ship in the future period. Moreover, the Dropout layer is introduced to describe the prediction uncertainty of the proposed model. Take the same trajectory sequence as the input, multiple groups of similar prediction results are obtained by randomly disabling several neural units of the CNN+LSTM network. Based on the statistical mean and variance of the prediction results, the prediction uncertainty of ship trajectory can be estimated. A dataset is created based on the open AIS data of confluence water in the coast of the United States. The input conditions are as follows: the input time is 60 min, the sam-pling frequency is 3 min, and the dropout parameter is 0.5. The results of the proposed model show that the error of trajectory prediction is 3.946 n mile for the next 60 min. The recognition accuracy of sailing intention is 87%. And the coverage rate of uncertainty estimation is 85.7%. Compared with other LSTM-based prediction methods, the trajectory prediction error of the proposed model is reduced by 31.6% when the ship's maneuverability changes. Furthermore, the proposed CNN+LSTM model has the ability of identifying ships' sailing intentions and estimating the prediction uncertainties, which is conducive to the development of intelligent navigation and intelligent maritime supervision technology.