基于稀疏模型的动态过程监测与故障分类研究

The main purpose of industrial process monitoring and fault diagnosis is to improve the quality of products and to ensure the process safety. However, industrial process model is complex and dynamic. Multivariate statistical monitoring model has poor interpretability and limited fault diagnosis information. These shortcomings make industrial process difficult to monitor. In this project, a dynamic process monitoring and fault classification method based on sparse model is proposed. The main contents include: (1) focusing on the sequential correlation between process variables, and proposing a multi-block monitoring strategy based on sequential information to strengthen the representation of dynamic characteristics, and to achieve the first step of sparsity; (2) using slow feature analysis to extract dynamic features in each block, constraining the L1-norm of load vector to achieve the second step of sparsity, building the sparse model so as to get the dynamic sparse representation of original variables, and monitoring dynamic process by constructing monitoring statistics based on sparse dynamic features; (3) utilizing sparse dynamic features to reduce the dimension of normal and faulted data, and adopting relevance vector machine algorithm to establish sparse multi-classification model to judge the fault types and to realize fault classification for dynamic process ; (4) verifying the monitoring performances of the proposed algorithm on Tennessee Eastman process and hot strip rolling process. The purpose of this project is to provide an interpretable, low complexity, and robust solution for dynamic industrial process monitoring and fault diagnosis.

工业过程监测与故障诊断是提高产品质量，保障生产安全的基础，然而其过程模型复杂、动态性强，且基于多元统计的监测模型可解释性差、故障诊断信息有限，导致了监测困难。对此，本课题提出基于稀疏模型的动态过程监测与故障分类方法，主要内容包括：（1）研究过程变量的时序相关性，提出基于时序信息的分块监测策略，加强动态特性在模型中的表达，实现模型第一步稀疏；（2）采用慢特征分析算法在每个模块内提取动态特征，通过约束负载向量的L1范数实现第二步稀疏，进而建立稀疏模型，得到原始变量的动态稀疏表达，并根据稀疏动态特征构造统计量，实现动态过程监测；（3）利用稀疏动态特征对正常/故障数据降维，采用相关向量机算法建立稀疏多分类模型，判断故障类别，实现动态过程故障分类；（4）在TE过程和带钢热连轧过程验证算法性能。本课题旨在为动态工业过程监测和故障诊断提供可解释的、复杂度低、鲁棒性强的解决方案。

工业过程监测与故障诊断是提高产品质量，保障生产安全的基础，然而其过程模型复杂、动态性强，且基于多元统计的监测模型可解释性差、故障诊断信息有限，导致了监测困难。对此，本课题以复杂动态工业过程为背景，探索动态过程的分布式监测技术，提出了一套基于稀疏模型的动态过程监测框架与稀疏故障分类方法，形成了动态过程监测与故障诊断的技术和原型软件。主要内容包括：（1）研究过程变量的时序相关性，提出基于时序信息的分块监测策略，构建了非平稳敏感子空间，加强动态特性在模型中的表达；（2）采用慢特征分析、协整分析算法在每个模块内提取动态特征，通过约束负载向量的L1范数实现第二步稀疏，进而建立稀疏模型，得到原始变量的动态稀疏表达，同时考虑特征的在线变化，根据稀疏动态特征构造统计量，实现动态过程监测；（3）提出了复杂工业过程质量预测模型，利用稀疏动态特征对正常/故障数据降维，采用相关向量机算法建立稀疏多分类模型，判断故障类别，实现动态过程故障分类；（4）在TE过程、带钢热连轧、乙酸乙烯酯生产等复杂工业过程中进行理论算法的实验验证。项目执行期间，在IEEE TII、ISA Transactions、Journal of Process Control、IEEE sensors Journal等国内外高水平学术期刊及中国自动化大会等国内外学术会议上发表论文19篇，其中SCI检索12篇，EI检索7篇；授权发明专利1项，登记软件著作权3项；项目负责人在项目执行期间晋升副教授职称，入选北京市科协青年人才托举工程，培养硕士2人，协助培养硕士3人。