融合深度学习的贝叶斯滤波综述  

A Survey on Bayesian Filtering With Deep Learning

作　　者：张文安[1,2] 林安迪 杨旭升 俞立[1,2] 杨小牛 ZHANG Wen-An;LIN An-Di;YANG Xu-Sheng;YU Li;YANG Xiao-Niu(College of Information Engineering,Zhejiang University of Technology,Hangzhou 310023;Zhejiang Provincial United Key Laboratory of Embedded Systems,Hangzhou 310023;National Key Laboratory of Electromagnetic Space Security,Jiaxing 314033)

机构地区：[1]浙江工业大学信息工程学院,杭州310023 [2]浙江省嵌入式系统联合重点实验室,杭州310023 [3]电磁空间安全全国重点实验室,嘉兴314033

出　　处：《自动化学报》2024年第8期1502-1516,共15页Acta Automatica Sinica

基　　金：国家自然科学基金(62173305);浙江省“尖兵”、“领雁”研发攻关计划(2022C03114);浙江省科技计划项目(2023C04032)资助。

摘　　要：当前动态系统呈现大型化、复杂化的趋势,基于贝叶斯滤波的动态系统状态估计遇到一系列新的挑战.随着深度学习在特征提取与模式识别等方面的优势与潜力不断显现,深度学习与传统贝叶斯滤波相结合的研究也随之兴起.为此,梳理了不同领域融合深度学习的贝叶斯滤波方法的应用案例,从中剖析不同类型动态系统下贝叶斯滤波存在的局限性和共性难题.在此基础上,总结了当前贝叶斯滤波存在的几类不确定性问题,以深度学习的视角将这些问题归纳为特征提取和参数辨识两大基本问题,进而介绍深度学习为贝叶斯滤波所提供的解决方案.其次,归纳整理了两类深度学习与贝叶斯滤波结合的具体方法,着重介绍了深度卡尔曼滤波和融合深度学习的自适应卡尔曼滤波.最后,综合考虑深度学习方法和贝叶斯滤波方法的优势,讨论了融合深度学习的贝叶斯滤波方法的开放问题和未来研究方向.As dynamic systems continue to exhibit a trend towards increased scale and complexity,the Bayesian filtering based state estimation for dynamic systems faces a series of new challenges.With the increasing prominence and new potential of deep learning in areas such as feature extraction and pattern recognition,research on combination of deep learning and classical Bayesian filtering is emerging.In this paper,we present a systematic review of application cases of Bayesian filtering methods that integrate deep learning in different domains,aiming to analyze the limitations and common challenges of Bayesian filtering in various types of dynamic systems.In view of this,we summarize several categories of uncertainty problems in the existing Bayesian filtering.From the perspective of deep learning,these problems are classified into two fundamental problems:Feature extraction and parameter identification.Furthermore,we introduce the solutions provided by deep learning for Bayesian filtering.Additionally,we categorize and organize two specific approaches that combine Bayesian filtering with deep learning,that is,deep Kalman filtering and adaptive Kalman filtering with deep learning.Finally,by considering the advantages of both deep learning and Bayesian filtering methods,we discuss open questions and future research directions for Bayesian filtering with deep learning.

关 键 词：深度学习 贝叶斯滤波 卡尔曼滤波 状态估计 状态空间模型 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]