Time-lags-aware distributed predictable feature analysis monitoring approach for large-scale dynamic processes

Abstract

In large-scale dynamic processes, variables often exhibit complex time-lag couplings and sequential redundancy that degrade model performance. However, traditional dynamic distributed approaches overlook the true lag structure during variable division and fail to effectively handle heterogeneous dynamic characteristics. In this article, by leveraging the capability of predictable feature analysis to construct diverse high-order dynamic models, a time-lags-aware distributed monitoring approach based on multi-block, multi-time-lag predictable feature analysis is proposed to address the aforementioned issues. First, a time-lags-aware variable division strategy based on sequential partial correlation is developed to eliminate sequential redundancy and divide the variables into subblocks according to their time lags. Subsequently, a subblock modeling method integrating predictable feature analysis with a customized time lag function is utilized to capture localized dynamic characteristics. Finally, Bayesian inference is introduced to synthesize the monitoring results from all subblocks, thereby determining the operating state of the current process. The superiority is validated on the Wastewater Treatment Process and Tennessee Eastman Process, where it achieves significantly improved fault detection rates compared to other methods.