Towards reliable control: Uncertainty-aware domain preserving stacked auto-encoder for data-driven modeling in large-scale industrial systems

Abstract

Online monitoring of operational states and quality indices in industrial processes is a vital source of information for enhancing production efficiency. This is particularly true for large-scale industrial systems, where existing methods often overlook the influence of uncertain information from abrupt operational fluctuations under time-varying processes and random noise during data collection and measurement. To address this issue, this paper proposes an uncertainty-aware key-domain-preserving stacked auto-encoder (UADP-SAE) model developed to capture the spatiotemporal distribution characteristics of dynamic operational changes in industrial systems. Based on this, an uncertainty quantification framework is designed to guide feature learning of the model by using uncertainty estimates. In addition, heteroscedastic uncertainty related to input noise is incorporated into the loss function, mitigating the adverse effects of high-uncertainty data on feature learning and enhancing the reliability of knowledge acquisition. Finally, the proposed method is validated on a real-world dataset from a large-scale industrial ironmaking system. Experimental results demonstrate that the proposed method outperforms traditional methods, achieving almost 10% improvement across multiple evaluation metrics for the prediction of three sintered ore quality indicators.