基于并发改进偏最小二乘的质量相关和过程相关的故障诊断
Quality-relevant and process-relevant fault diagnosis with concurrent modified partial least squares
摘要点击 116  全文点击 51  投稿时间:2020-06-11  修订日期:2020-09-24
查看全文  查看/发表评论  下载PDF阅读器
DOI编号  10.7641/CTA.2020.00345
  2021,38(3):318-328
中文关键词  过程控制  主元分析  故障诊断  偏最小二乘  安全
英文关键词  process control  principal component analysis  fault diagnosis  partial least squares  safety
基金项目  国家自然科学基金项目(61673387, 61833016), 陕西省自然科学基金项目(2020JM–356)资助.
作者单位E-mail
李强 火箭军工程大学 leeqang@yeah.net 
孔祥玉 火箭军工程大学 xiangyukong01@163.com 
罗家宇 火箭军工程大学  
解建 火箭军工程大学  
中文摘要
      并发潜结构投影(CPLS)与传统贡献图法是多元统计过程监控中常用的故障检测与诊断方法. 过程监控通 常要求监测的时效性与诊断的准确性, 然而, 由于CPLS计算复杂以及传统贡献图诊断结果易受初始贡献较大的变 量影响, 因此它们反馈的监控结果可能并不准确. 针对上述问题分别提出一种并发改进偏最小二乘(CMPLS)方法和 新的相对贡献图法(NRC). 首先, CMPLS将输入和输出数据同时投影到与过程相关或质量相关的多个子空间, 在相 应子空间分别构造适用于各种故障报警的监测指标进行过程监测; 然后, 结合所提出的NRC进行故障识别. 所提方 法对过程故障实现全面监测的同时避免了过多的迭代过程, 并消除了过程变量中对检测指标初始贡献较大变量的 影响. 最后利用数值仿真和田纳西伊士曼过程验证了所提方法的有效性.
英文摘要
      The concurrent projection to latent structures (CPLS) and conventional contribution plots are often used as fault detection and diagnosis approaches for multivariate statistical process monitoring. Process monitoring usually requires the timeliness of monitoring and the accuracy of diagnosis, but the complex calculation of CPLS and the diagnosis results of conventional contribution plots are susceptible to the variables with large initial contributions, thereby their feedback monitoring results may not be accurate. To solve the above problems, a concurrent modified-PLS (CMPLS) method and a new relative contribution plots (NRC) method are proposed respectively. Originally, CMPLS projects the input and output data to multiple subspaces related to the process or quality at the same time, and respectively constructs monitoring indicators suitable for various fault alarms in the corresponding subspaces to monitor the process; then combined with the proposed NRC to identify the faults. The proposed methods realize comprehensive monitoring of process faults while avoiding too many iterative processes, and eliminates the influence of variables which larger initial contribution in process variables to the detection index. Finally, numerical simulation and Tennessee Eastman process are used to verify the efficiency of the proposed methods.