Soft-sensing model development using PLSR-based dynamic extreme learning machine with an enhanced hidden layer

Soft sensors have been widely used as online instrument measurements for the key process variables of industrial processes. In this paper, a novel robust soft sensor model for predicting the key process variables is proposed. The proposed soft sensor model integrates an enhanced hidden layer based dynamic extreme learning machine (EHLDELM) with the partial least-square regression (PLSR). The traditional extreme learning machine with a static structure cannot well deal with the dynamic feature of the process data, so a dynamic strategy is adopted. Additionally, a special linear hidden layer node is added in the dynamic extreme learning machine to further enhance the performance. Then, the partial least-square method is utilized to deal with the collinearity problem. Finally, an optimal model between the hidden layer and the output layer is obtained. Thus, a novel robust nonlinear soft sensor model integrated EHLDELM with PLSR (PLSR-EHLDELM) is proposed. As a case study, the proposed PLSR-EHLDELM model is demonstrated through an application to the Tennessee Eastman process (TEP) for estimation of the key process variable. Compared with the other four models of ELM, PLSR, Kernel-based PLS, and PLS-ELM, the proposed PLSR-EHLDELM model achieves higher accuracy.