A robust data-driven approach for modeling industrial systems with non-stationary and sparsely sampled data streams

Abstract

Data-driven modeling has been widely applied to estimate quality variables in industrial processes, while several practical challenges hinder their applications. This paper is concerned with data-driven modeling for industrial plants with non-stationary and sparsely sampled data streams. To avoid overfitting from data scarcity, a Bayesian linear model is preferred and an associated identification algorithm based on variational inference is developed. The key contribution of this work lies in the development of an adaptation mechanism using streaming variational Bayes with power priors, enabling model identification and adaptation to non-stationary and sparsely sampled data within a full Bayesian framework. To enhance robustness against outliers in streaming batches, Student’s-t distribution is used to account for noise. Furthermore, a posterior predictive distribution is approximately derived, allowing the model to provide not only a point estimate but also the associated predictive uncertainty. The effectiveness of the proposed method is validated through a numerical example and an industrial application.