Development of a Data-Driven Soft Sensor for Multivariate Chemical Processes Using Concordance Correlation Coefficient Subsets Integrated with Parallel Inverse-Free Extreme Learning Machine

Abstract

Nonlinearity, complexity, and technological limitations are causes of troublesome measurements in multivariate chemical processes. In order to deal with these problems, a soft sensor based on concordance correlation coefficient subsets integrated with parallel inverse-free extreme learning machine (CCCS-PIFELM) is proposed for multivariate chemical processes. In comparison to the forward propagation architecture of neural network with a single hidden layer, i.e., a traditional extreme learning machine (ELM), the CCCS-PIFELM approach has two notable points. Firstly, there are two subsets obtained through the concordance correlation coefficient (CCC) values between input and output variables. Hence, impacts of input variables on output variables can be assessed. Secondly, an inverse-free algorithm is used to reduce the computational load. In the evaluation of the prediction performance, the Tennessee Eastman (TE) benchmark process is employed as a case study to develop the CCCS-PIFELM approach for predicting product compositions. According to the simulation results, the proposed CCCS-PIFELM approach can obtain higher prediction accuracy compared to traditional approaches.