文献类型：期刊文献

英文题名：Interval-Valued-Based Stacked Attention Autoencoder Model for Process Monitoring and Fault Diagnosis of Nonlinear Uncertain Systems

作者：Wu, Qiqi[1];Yan, Xuefeng[1]

机构：[1]East China Univ Sci & Technol, Key Lab Smart Mfg Energy Chem Proc, Minist Educ, Shanghai 200237, Peoples R China

年份：2023

卷号：72

外文期刊名：IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT

收录：;EI(收录号：20224813194298);WOS:【SCI-EXPANDED(收录号:WOS:000915866600003)】；

基金：This work was supported in part by the National KeyResearch and Development Program of China under Grant 2021YFC2101100 and in part by the National Natural Science Foundation of China under Grant 21878081. The Associate Editor coordinating the review process was Dr. Md. Moinul Hossain

语种：英文

外文关键词：Feature extraction; Data models; Process monitoring; Principal component analysis; Data mining; Fault diagnosis; Computational modeling; Attention mechanism (AM); fault-related variables; Jensen-Shannon (JS) divergence; stacked autoencoder (SAE); Tennessee Eastman (TE) process

摘要：The measurement data fluctuate up and down within an interval centered on the true value due to disturbances and noise which are zero mean. Meanwhile, traditional industrial process monitoring algorithms are mostly based on normal data for modeling and rarely consider fault information. As a result, the variables involved in the modeling process may contain information irrelevant to the fault, thereby leading to the degradation of monitoring performance. On the basis of the above considerations and to avoid the occurrence of major safety accidents, this study proposes a stacked attention autoencoder (SAAE) monitoring model in view of the upper and lower bounds of the interval and fault-related variables. Based on the viewpoint that some sampled variables' distribution will change after the fault occurs, Jensen-Shannon (JS) divergence is used as an indicator to measure the difference, and the variables with significant changes before and after the fault are screened out. Subsequently, an attention mechanism (AM) is introduced in the process of training stacked AE (SAE). Thus, the features with a strong correlation with the just-screened fault-related variables have a larger weight. In other words, the model focuses more on the fault-related features. This method not only reduces the influence of uncertainty, but also uses historical fault data to pick out fault-related information. This study demonstrates the performance of the algorithm through the Tennessee Eastman (TE) process.