Neural networks-based hybrid beneficial variable selection and modeling for soft sensing.

Variable selection plays an important role in soft sensor development. Either including redundant variables or missing important variables can degrade the modeling performance, affecting practical industrial applications. This paper proposes a novel neural networks-based hybrid beneficial variable selection (HBVS) and modeling method for effective soft sensing. First, irrelevant variables are removed through evaluating mutual information (MI) between all candidate variables and the quality variable. Second, proxy variables are introduced and a hidden gain-based evaluation method is employed to temporarily sort variables according to their significance, which facilitates to make use of process knowledge. Then, false discovery rate is employed to identify the model consistency, through which beneficial variables are determined. The proposed soft sensor development method is tested on a penicillin simulation process, and two actual industrial processes, including an oil refining process and an actual penicillin production process. Comparisons to state-of-the-art existing methods verify the effectiveness and superiority of the proposed method. • A neural networks-based variable selection and modeling method is proposed. • It is an end-to-end feature selection method for nonlinear soft sensor. • Three application examples are provided. • Comparisons to state-of-the-arts show effectiveness of the proposed method. [ABSTRACT FROM AUTHOR]