Time neighborhood preserving auto-regressive model for industrial process monitoring

Abstract The mandatory requirement of ensuring plant safety as well as sustainable operation keeps motivating novel approaches for trustfully detecting anomalies or faults during production. Recently, a unified modeling framework which extracts latent variables (LVs) with explicit interpretation to time-serial variation, is becoming quite prevalent in the circle of data-driven process monitoring. In order to faithfully characterize the underlying time-serial variation in sequentially measured samples, a novel analytic algorithm called time neighborhood preserving auto-regressive model (TNPARM) is proposed. The main advantages of TNPARM benefit from the following two considerations: (1) each sample and its pasts stacked as an augmented vector are reconstructed from corresponding time neighborhood, so that time sequential interdependencies can be structured and then preserved through projection; (2) an auto-regressive model is sought to represent the auto-correlation between individual LVs and one-step ahead prediction thereof, so that the remaining dynamics in reconstruction residuals could be further exploited. A novel optimization problem is designed and an iterative algorithm is provided to realize the two concerns derived for TNPARM. Finally, the superiority of TNPARM in process monitoring is validated through a series of comparative experiments on two industrial-scale examples, i.e., the Tennessee Eastman benchmark process and a real-world multiphase flow facility.