Ensemble Quality-Aware Slow Feature Analysis for decentralized dynamic process monitoring

Abstract

Slow Feature Analysis (SFA) has gained prominence in process monitoring due to its capability to capture inertial features in industrial systems. However, traditional SFA methods are predominantly unsupervised and often neglect output quality, limiting their effectiveness in large-scale, complex systems. To address these limitations, this paper introduces the Ensemble Quality-Aware Slow Feature Analysis (EQASFA) framework, which maximizes the correlation between quality variables and slow features. This decentralized monitoring framework generates fine-grained submodels by: (i) constructing a diverse set of submodels through different variable combinations, and (ii) selecting base submodels with the lowest false alarm rate on the validation dataset. The selection process utilizes a divisive hierarchical clustering algorithm, where probabilistic similarity is quantified using symmetric Kullback–Leibler divergence. In addition, novel static and dynamic metrics, derived from Bayesian inference, are proposed to distinguish routine operational fluctuations from significant anomalies. The performance of the EQASFA framework is validated through two benchmark case studies: the Tennessee Eastman process and a wastewater treatment process.