欢迎访问《兵工学报》官方网站,今天是

兵工学报 ›› 2025, Vol. 46 ›› Issue (6): 240578-.doi: 10.12382/bgxb.2024.0578

• • 上一篇    下一篇

基于深度学习的联合等式状态约束辨识与递推滤波

秦月梅1,*(), 陈重1, 杨衍波2,3, 李淑英1   

  1. 1 西安邮电大学 自动化学院, 陕西 西安 710121
    2 西北工业大学 自动化学院, 陕西 西安 710129
    3 信息融合技术教育部重点实验室, 陕西 西安 710129
  • 收稿日期:2024-07-12 上线日期:2025-06-28
  • 通讯作者:
  • 基金资助:
    国家自然科学基金项目(61903299); 陕西省自然科学基金项目(2024JC-YBMS-457)

Joint State Equality Constraint Identification and Recursive Filtering Based on Deep Learning

QIN Yuemei1,*(), CHEN Zhong1, YANG Yanbo2,3, LI Shuying1   

  1. 1 School of Automation, Xi’an University of Posts & Telecommunications, Xi’an 710121, Shaanxi, China
    2 School of Automation, Northwestern Polytechnical University, Xi’an 710129, Shaanxi, China
    3 Key Laboratory of Information Fusion Technology, Ministry of Education, Xi’an 710129, Shaanxi, China
  • Received:2024-07-12 Online:2025-06-28

摘要:

针对等式约束跟踪系统中多约束并存且当前约束信息不确定下状态估计问题,提出基于深度学习的联合等式状态约束辨识与递推滤波算法。利用门控循环单元构建约束判别网络,借助雷达量测实现当前时刻等式状态约束的在线辨识;在递推滤波框架下基于级联门控循环单元构建增益学习网络,实现概率模型与数据学习联合驱动的目标状态自适应估计;通过滤波投影联合约束判别网络辨识的约束信息和增益学习网络输出的状态估计,获得满足当前时刻等式状态约束的高精度目标状态估计。典型多道路目标跟踪实验结果表明:新算法相比卡尔曼滤波、交互式多模型(基于不同运动模型构建模式集/不同等式状态约束构建模式集)和KalmanNet等算法,在不同量测噪声水平下具有更高的估计精度和更好的鲁棒性。

关键词: 目标跟踪, 等式约束, 线性随机系统, 状态估计, 深度学习

Abstract:

The state estimation problem of equality constraint-based target tracking where multiple constraints coexist and the current constraint information is unknown is presented.A joint state equality constraint identification and recursive filtering based on deep learning algorithm is proposed.The gated recurrent units are used to construct a constraint discriminant network,and the current state constraint is identified online by using the radar measurement.A filtering gain learning network based on cascaded gated recurrent units is built in the framework of recursive filtering to adaptively estimate the target state with the help of joint probabilistic modeling and data learning.The final high-precision filtered estimate which meets the real state equality constraint at current sampling instant is obtained based on filtering projection,which combines the state estimate obtained by the gain learning network with the state equality constraint identified by the constraint discriminant network.Experimental results of multi-target tracking example demonstrate that the proposed algorithm outperforms Kalman filtering,interacting multiple model (based on different motion modes/different state equality constraints) and KalmanNet algorithm in terms of estimation accuracy and robustness,with different levels of measurement noises.

Key words: target tracking, equality constraint, linear stochastic system, state estimation, deep learning

中图分类号: