Adaptive ranking based ensemble learning of Gaussian process regression models for quality-related variable prediction in process industries

Liu, Yiqi and Huang, Daoping and Liu, Bin and Feng, Qiang and Cai, Baoping (2021) Adaptive ranking based ensemble learning of Gaussian process regression models for quality-related variable prediction in process industries. Applied Soft Computing, 101. 107060. ISSN 1568-4946 (https://doi.org/10.1016/j.asoc.2020.107060)

[thumbnail of Liu-etal-ASC-2020-Adaptive-ranking-based-ensemble-learning]
Preview
Text. Filename: Liu_etal_ASC_2020_Adaptive_ranking_based_ensemble_learning.pdf
Accepted Author Manuscript
License: Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 logo

Download (1MB)| Preview

Abstract

The proper monitoring of quality-related but hard-to-measure variables is currently one of the bottlenecks limiting the safe and efficient operations of industrial processes. This paper proposes a novel ensemble learning algorithm by coordinating global and local Gaussian process regression (GPR) models, and this algorithm is able to capture global and local process behaviours for accurate prediction and timely process monitoring. To further address the deterioration in predictions when using the off-line training and online testing strategy, this paper proposes an adaptive ranking strategy to perform ensemble learning for the sub-GPR models. In this adaptive strategy, we use the moving-window technique to rank and select several of the best sub-model predictions and then average them together to make the final predictions. Last but not least, the least absolute shrinkage and selection operator (Lasso) works together with factor analysis (FA) in a two-step variable selection method to remove under-correlated model input variables in the first stage and to compress over-correlated model input variables in the second stage. The proposed prediction model is validated in two real wastewater treatment plants (WWTPs) with stationary and nonstationary behaviours. The results show that the proposed methodology achieves better performance than other standard methods in the context of their predictions of quality-related variables.