模型误差影响下基于CNN+BiLSTM神经网络的非圆信号目标直接跟踪算法  

Direct Tracking for a Non-Circular Source Based on CNN+BiLSTM Neural Network in the Presence of Modeling Errors

作　　者：尹洁昕 王鼎 杨欣[3] 杨宾 YIN Jie-xin;WANG Ding;YANG Xin;YANG Bin(Institute of Information System Engineering,PLA Strategic Support Force Information Engineering University,Zhengzhou,Henan 450001,China;National Digital Switching System Engineering and Technology Research Center,Zhengzhou,Henan 450002,China;Institute of Computer and Artificial Intelligence,Zhengzhou University,Zhengzhou,Henan 450001,China)

机构地区：[1]中国人民解放军战略支援部队信息工程大学信息系统工程学院,河南郑州450001 [2]国家数字交换系统工程技术研究中心,河南郑州450002 [3]郑州大学计算机与人工智能学院,河南郑州450001

出　　处：《电子学报》2024年第4期1315-1329,共15页Acta Electronica Sinica

基　　金：国家自然科学基金(No.61901526,No.62171469,No.62071029);军事科技领域青年人才托举工程(No.2022-JCJQ-QT-028);全军共用信息系统装备预先研究专用技术项目(No.315087701)。

摘　　要：针对运动观测阵列状态误差与接收频率抖动同时影响下的非圆信号无源跟踪问题,提出了一种基于卷积神经网络(Convolutional Neural Network,CNN)+双向长短时记忆网络(Bi-directional Long Short-Term Memory,BiL⁃STM)的直接跟踪算法.该算法首先利用多运动观测阵列信号各频带间的相关性与辐射源信号的非圆特性,建立模型误差影响下的扩展多站观测矢量;接着利用多个观测时隙内扩展多站观测矢量的信号子空间构造空时特征输入序列;然后设计基于CNN与BiLSTM混合神经网络的直接跟踪模型,通过训练实现对非圆目标的轨迹矢量直接估计.本文算法是从原始数据信号子空间中估计轨迹矢量的直接跟踪模式,相比传统“观测参数估计+滤波轨迹跟踪”的两步估计模式,具有更高的估计精度.由于本文算法在神经网络训练过程中学习到模型误差的信息,因此能够实现对多种误差的校正.仿真结果表明,本文算法较传统两步跟踪算法与现有直接跟踪算法均具有更高的轨迹估计精度,能够明显提升模型误差影响下多站协同跟踪的鲁棒性.In order to track a non-circular source target under the influence of both the state errors of moving observer arrays and frequency jitter deviations,this paper proposes a direct tracking algorithm based on the CNN(Convolutional Neu⁃ral Network)+BiLSTM(Bi-directional Long Short-Term Memory)neural network.The proposed algorithm first exploits the correlation between the received array signals of multiple moving observers in various frequency bands and the non-circular property of the radiation source signals.Therefore,an extended multi-station observation vector is established in the presence of modeling errors.Then,the non-circular signal subspaces of the extended multi-station observation vectors within multiple observing timeslots are used to form the spatiotemporal feature input sequence.Subsequently,the direct tracking model for a non-circular source target based on CNN and BiLSTM neural network is designed.After training this neural network,the tra⁃jectory vector for a non-circular source target in several timeslots can be directly determined.Because the proposed algorithm estimates the target trajectory vector directly from the subspaces of the original array signals,it has higher estimation accuracy compared to the traditional two-step tracking methods which extract measurement parameters and then estimate the trajectory from them.As the information of modeling errors can be learned from training the neural network,this algorithm can achieve the calibration for modeling errors.The simulation results show that the proposed algorithm has higher trajectory estimation accuracy compared to the traditional two-step tracking algorithm and existing direct tracking algorithm.It significantly im⁃proves the robustness of multi-station collaborative tracking under the influence of modeling errors.

关 键 词：直接跟踪 非圆信号 模型误差 卷积神经网络 双向长短时记忆网络 

分 类 号：TN911.7[电子电信—通信与信息系统]