目前已有许多工作将Transformer运用到时间序列预测相关任务. 然而, 相比其他时间序列, 运动轨迹数据存在运动学的不确定性, 没有明显的周期特性. 为了降低噪声干扰, 增强趋势建模, 本文在Transformer架构的基础上, 提出一种基于时频域信息融合和多尺度对抗训练的目标轨迹预测方法. 将小波分解嵌入网络模型, 实现时频域自适应滤波; 并与时域注意力进行融合, 能够更有效地对观测轨迹的长期趋势特性进行编码. 并设计了一个全卷积判别器, 通过对抗训练学习序列的多尺度短期微运动表示, 进一步提高预测精度. 本文建立了一个包括2维船舶轨迹和3维飞行器轨迹的轨迹预测数据集DT作为基准, 并在此与Transformer、LogTrans、Informer等模型进行对比实验. 实验结果表明本文的方法在中长期轨迹预测任务上优于其他模型.
Many studies apply Transformer to time series prediction tasks. However, compared with other time series, motion trajectory data has kinematic uncertainty without obvious periodicity. To reduce noise interference and enhance trend modeling, this study proposes a target trajectory prediction method based on time-frequency domain information fusion and multi-scale adversarial training based on Transformer architecture. The wavelet decomposition is embedded into the network model to realize the adaptive filtering in the time-frequency domain, and then time-domain attention is integrated to encode the long-term trend characteristics of the observed trajectory more effectively. Meanwhile, the study designs a full convolution discriminator to further improve the prediction accuracy by learning multi-scale short-term micro motion representation of the sequence through adversarial training. A trajectory prediction dataset DT including 2D ship trajectory and 3D aircraft trajectory is established as a benchmark, and comparative experiments with Transformer, LogTrans, Informer, and other models are conducted. Experiment results show that the proposed method is superior to other models in the tasks of medium and long-term trajectory prediction.