基于改进注意力机制的自适应航迹预测方法  

Adaptive Trajectory Prediction Method Based on Improved Attention Mechanism

作　　者：黄权印 蔡益朝[1] 李浩[1] 唐晓[1] 王辰洋 HUANG Quanyin;CAI Yichao;LI Hao;TANG Xiao;WANG Chenyang(Air Force Early Warning Academy,Wuhan 430000,Hubei,China)

机构地区：[1]空军预警学院,湖北武汉430000

出　　处：《空天防御》2024年第3期94-101,共8页Air & Space Defense

基　　金：国家自然科学基金青年(61502522);国家社科基金重点(2022-SKJJ-B-056);湖北省自科基金面上项目(2019CFC897)。

摘　　要：针对现有循环神经网络在解决目标航迹预测问题中容易存在训练过拟合、预测精度不高、泛化能力差以及适应性不强的问题,提出了一种基于改进注意力机制和门控循环单元(GRU)的目标航迹预测方法。该方法通过早停法来自动终止网络训练过程,防止训练过拟合;通过模型检查点函数保存网络训练中的最优网络参数;通过把注意力机制引入GRU网络中,对轨迹特征赋予不同的权重来聚焦重点航迹信息,提高网络的预测性能。最后,通过仿真实验证明,该方法能够有效提升循环神经网络的预测精度、泛化性及适应性。The existing recurrent neural networks are subject to training overfitting,low prediction accuracy,poor generalization ability,and weak adaptability in solving target trajectory prediction.A target trajectory prediction method using an improved attention mechanism and Gated Recurrent Unit(GRU)was proposed,which could automatically terminate the network training process through an early stopping method to prevent overfitting during training.It saved the optimal network parameters during network training through the model checkpoint function.By introducing an attention mechanism into the GRU network and assigning different weights to trajectory features to focus on key trajectory information,the predictive performance of the network was optimized Finally,simulation experiments results show that the proposed method effectively improves the prediction accuracy,generalization,and adaptability of recurrent neural networks.

关 键 词：航迹预测 注意力机制 早停法 循环神经网络 门控循环单元 

分 类 号：TN953[电子电信—信号与信息处理]