Soft Sensor Development Using Improved Whale Optimization and Regularization-Based Functional Link Neural Network
Ye Tian, Yan-Lin He, Qunxiong Zhu
Abstract
Recently, data-driven soft sensor has been a popular research focus in the field of process system engineering. Modern industrial processes tend to be large scale, highly complicated, and nonlinear. As a result, process data gradually become high-dimensional. Therefore, it is difficult to achieve acceptable modeling accuracy using basic data-driven methods. To handle this limitation, a novel data-driven model using improved whale optimization and regularization-based functional link neural network (FLNN) is proposed. In the proposed model, regularization is first used to overcome the problems of structure risk and overfitting during the training phase of FLNN, thereby improving its ability to deal with the complex process data; to simplify the calculation, a radial basis function (RBF)-based kernel is selected to reconstruct the expanded inputs; meanwhile, an improved whale optimization algorithm (WOA) is utilized to optimize the parameters of the regularization and RBF kernel. Finally, novel regularized FLNN based on WOA and RBF kernel (WOA-RBFRFLNN) can be developed. To verify the modeling performance of WOA-RBFRFLNN, a case study on the purified terephthalic acid (PTA) industrial process is conducted. Simulation results show that the presented WOA-RBFRFLNN model can achieve high accuracy, indicating that the feasibility and effectiveness of the proposed WOA-RBFRFLNN are confirmed.