Joint time-serial variation analysis for fault monitoring of chemical processes

• Published in: Process safety and environmental protection Vol. 196; p. 106867
• Main Authors: Hu, Guowei, Tong, Chudong, Zeng, Jia, Luo, Lijia
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2025
• Subjects: Dynamic latent variable; Fault monitoring; Feature extraction; Time-serial variation; Time-serial variation; Feature extraction; Fault monitoring; Dynamic latent variable
• ISSN: 0957-5820
• DOI: 10.1016/j.psep.2025.106867

As the classical analytics in the field of data-driven fault monitoring, time neighborhood preserving embedding (TNPE), dynamic inner canonical correlation analysis (DiCCA), and slow feature analysis (SFA) provided three different choices for characterizing time-serial variation inherent in sequential samples. Considering the unsupervised nature of these three algorithms as well as their variants, it could be more appropriate to jointly exploit time-serial variation from multiple perspectives in a comprehensive manner. This recognition then motivates us to propose a novel dynamic modeling algorithm titled as joint time-serial variation analysis (JTSVA) for fault monitoring. The proposed JTSVA aims to extract dynamic latent variables (DLVs) with respect to a joint integration of time-manifold embedding, latent auto-regressive, and slow-varying capabilities own by TNPE, DiCCA and SFA, respectively. Furthermore, an additional orthogonality constraint is further assigned to the problem definition of JTSVA, so that the extracted DLVs could have enhanced discriminant in uncovering valuable information for satisfactory fault monitoring performance. Finally, the superiority of JTSVA in fault monitoring, in terms of false alarm rate and fault detection rate, is validated through comparative experiments on two industrial-scale examples, i.e., the Tennessee Eastman benchmark process and a real-world multiphase flow facility.