自动化学报2024，Vol.50Issue(8)：1502-1516,15.DOI:10.16383/j.aas.c230457
融合深度学习的贝叶斯滤波综述
A Survey on Bayesian Filtering With Deep Learning
摘要
Abstract
As dynamic systems continue to exhibit a trend towards increased scale and complexity,the Bayesian fil-tering 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 combina-tion 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 iden-tification.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Key words
Deep learning/Bayesian filtering/Kalman filtering/state estimation/state-space model引用本文复制引用
张文安,林安迪,杨旭升,俞立,杨小牛..融合深度学习的贝叶斯滤波综述[J].自动化学报,2024,50(8):1502-1516,15.基金项目
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Supported by National Natural Science Foundation of China(62173305),"Pioneer","Leading Goose"Research and Develop-ment Program of Zhejiang Province(2022C03114),and Develop-ment Program of Zhejiang Province(2023C04032) （62173305）
