基于残差修正CNN-BiLSTM的空中目标航迹短期预测算法  

Short-term prediction algorithm of air target track based on residual correction CNN-BiLSTM

作　　者：王硕[1] 吴楠[1] 黄洁[1] 王建涛 WANG Shuo;WU Nan;HUANG Jie;WANG Jiantao(University of Information Engineering,Zhengzhou 450001,China)

机构地区：[1]信息工程大学,河南郑州450001

出　　处：《指挥控制与仿真》2024年第1期55-63,共9页Command Control & Simulation

摘　　要：针对因深度学习自身局限性和递归预测策略产生的累积误差,导致航迹预测精度不高的问题,提出了一种基于残差修正CNN-BiLSTM的空中目标航迹短期预测算法。首先,引入卷积模块用于提取航迹数据之中具有潜在关联的空间位置特征,利用双向长短时记忆网络提取航迹数据中的时序特征,并实现对空中目标的实时单步预测和多步超前预测;其次,引入整合移动平均自回归为残差修正模型,对实时单步预测产生的残差建模,计算混合神经网络模型多步超前预测时的残差值;最后,将混合神经网络模型和残差修正模型的输出结果进行融合,得到最终的航迹预测值。实验结果表明,该算法大大降低了神经网络因自身局限性产生的误差和因递归策略预测产生的累积误差,能够显著提高空中目标航迹短期预测的精度。To solve the problem of low track prediction accuracy caused by the limitations of deep learning and the cumulative error generated by recursive prediction strategies,a short-term prediction algorithm for air target tracks based on residual correction CNN-BiLSTM was proposed.Firstly,a convolution module was introduced to extract potentially associated spatial location features from the track data,and a bidirectional long and short time memory network was used to extract temporal features from the track data,achieving real-time one-step prediction and multi-step advance prediction of air targets.Then,the integrated moving average autoregression was introduced as a residual correction model to model the residual generated by real-time one-step prediction,and the residual value of the hybrid neural network model for multi-step advanced prediction is calculated.Finally,the output results of the hybrid neural network model and the residual correction model are fused to obtain the final trajectory prediction value.Experiment results proved that the algorithm can significantly improve the accuracy of short-term prediction of airborne target tracks.

关 键 词：残差修正 CNN-BiLSTM 短期预测 ARIMA 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程] E926.4[自动化与计算机技术—控制科学与工程]