Multi-subspace quality-aware slow feature extraction for decentralized process monitoring

Abstract

The focus on process safety and product quality has propelled the adoption of sophisticated data-driven methodologies, with multivariate statistical process monitoring (MSPM) becoming a cornerstone in process industries. However, as these industries expand and generate increasingly complex datasets, conventional MSPM frameworks often fail to capture intricate process dynamics, complicating fault detection and diagnosis. Additionally, most MSPM approaches do not consider the moderating influence of closed-loop controllers on abnormal conditions, frequently leading to misclassification of operational transitions as faults. To address these challenges, this paper presents a Multi-Subspace Quality-Aware Slow Feature Analysis (MQASFA) method for concurrently monitoring operational deviations and anomalous behavior. The MQASFA framework employs a multi-subspace strategy using symmetric Kullback-Leibler divergence to aggregate correlated variables from a probabilistic standpoint. A divisive hierarchical clustering algorithm is applied to integrate variable blocks, reducing computational complexity and redundancy while preserving essential local process information. A quality-aware slow feature analysis submodel is subsequently deployed for decentralized, quality-focused monitoring within each subspace. To differentiate routine operational variations from significant anomalies, novel static and dynamic metrics derived from Support Vector Data Description are introduced. The efficacy of the proposed methodology is validated through applications to the Tennessee Eastman process and a wastewater treatment benchmark.