引用本文:李灵,刘自鹏,王雅琳,刘述,潘卓夫,李磊.ELM-AAE驱动的工业过程故障诊断与故障深度估计[J].控制理论与应用,2026,43(5):1133~1141.[点击复制]
LI Ling,LIU Zi-peng,WANG Ya-lin,LIU Shu,PAN Zhuo-fu,LI Lei.ELM-AAE driven industrial process fault diagnosis and fault depth estimation[J].Control Theory & Applications,2026,43(5):1133~1141.[点击复制]
ELM-AAE驱动的工业过程故障诊断与故障深度估计
ELM-AAE driven industrial process fault diagnosis and fault depth estimation
摘要点击 332  全文点击 11  投稿时间：2024-02-17  修订日期：2025-11-22
查看全文  查看/发表评论  下载PDF阅读器   HTML
DOI编号  10.7641/CTA.2025.40103
  2026,43(5):1133-1141
中文关键词  对抗自编码器  故障诊断  互信息  故障估计  梯度下降算法
英文关键词  adversarial autoencoder  fault diagnosis  mutual information  fault estimation  gradient descent algorithm
基金项目  国家自然科学基金项目(62103063, 62303494), 湖南省教育厅科学研究项目(25A0189), 湖南省自然科学基金项目(2022JJ40510, 2024JJ6722), 湘江 实验室重大项目(23XJ01009)资助.
作者单位E-mail
李灵 长沙理工大学人工智能学院 lilings@csust.edu.cn 
刘自鹏 长沙理工大学人工智能学院  
王雅琳 中南大学自动化学院  
刘述 杭州智元研究院有限公司  
潘卓夫* 湖南工商大学微电子与物理学院 joffpan_ai@outlook.com 
李磊 长沙理工大学人工智能学院  
中文摘要
      对抗自编码器(AAE)模型训练时, 样本与特征表示间的互信息衰减现象, 不仅影响模型的故障检测性能, 而 且难以直接用于故障估计. 为此, 本文设计了一种基于逐层扩展互信息对抗自编码器的故障检测策略, 通过合并隐 层特征与原始数据, 显性引入特征空间与每层神经网络的互信息, 实现了信息的高效整合及特征与输入样本间的最 大化相关性; 并在此基础上, 开发了一种梯度下降反馈算法, 设计了基于深度学习的故障估计策略, 有效避免了故障 估计的复杂机理建模与大规模系统参数辨识; 最后, 通过连续搅拌釜式加热器验证了所提方法的有效性和优越性.
英文摘要
      The mutual information decay between samples and feature representations during the training of the adversarial autoencoder (AAE) model not only affects the fault detection performance, but also makes it difficult to be used directly for fault estimation. To this end, this paper designs a fault detection strategy based on extended layer-by-layer mutual information adversarial autoencoder (ELM-AAE), which achieves efficient integration of information and maximized correlation between features and input samples by merging the hidden layer features with the original data and explicitly introducing mutual information between the feature space and each layer of the neural network. A gradient descent feedback algorithm is then developed to construct a fault estimation model based on deep learning, which effectively avoids the complex mechanism modeling and large-scale system parameter identification for fault estimation. Finally, the validity and superiority of the proposed method was verified by continuous stirring tank heater (CSTH).