Fault diagnosis of chemical processes based on joint recurrence quantification analysis

• Published in: Computers & chemical engineering Vol. 155; p. 107549
• Main Authors: Ziaei-Halimejani, Hooman, Nazemzadeh, Nima, Zarghami, Reza, Gernaey, Krist V., Andersson, Martin Peter, Mansouri, Seyed Soheil, Mostoufi, Navid
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.12.2021
• Subjects: DBSCAN; Fault diagnosis; Joint recurrence plot; Missing data; Unsupervised learning; Fault diagnosis; Missing data; DBSCAN; Joint recurrence plot; Unsupervised learning
• DOI: 10.1016/j.compchemeng.2021.107549
• ISSN: 0098-1354

•An unsupervised JRQA-based fault diagnosis scheme was proposed for handling of missing data.•Three nonlinear, unstable and nonstationary multivariate chemical process were analyzed.•Compatibility and impact of various imputation methods were assessed.•JRQA showed lower sensitivity and higher robustness in case of missing data.•JRQA method can be tuned for different frequency and scale of missing data. An unsupervised learning method is developed for fault detection and diagnosis with missing data for chemical processes based on the multivariate extension of joint recurrence quantification analysis (JRQA) and clustering. The application of the proposed method is assessed in the presence and absence of imputation methods. To provide a comprehensive scheme, three different processes were utilized including, silica particle flocculation (SFP) as an unstable batch process, a chemical looping combustion (CLC) process, and the Tennessee Eastman process (TEP) as the control system design benchmark. The application of the developed method demonstrated that the JRQA method has the best performance in fault diagnosis of the complete dataset in all three processes compared to previously developed methods. Moreover, in the case of missing data, the sensitivity of the results can be adjusted by changing the length of the sub-series. The sensitivity of the proposed method is 33% lower for SFP, 30% for CLC and 32% for TEP, compared to the probabilistic kernel principal components analysis (PKPCA)-based method. [Display omitted]