Nonlinear soft sensor development for industrial thickeners using domain transfer functional-link neural network

Abstract

The thickener is used to provide slurries with a stable and satisfactory concentration in the ore dressing plant. To efficiently control an industrial thickener, a soft sensor model should be built first to predict the underflow concentration. In industrial sites, it is usually expensive and time-consuming to collect sufficient high-quality data to develop a data-driven model. In this work, a nonlinear regression method for transfer learning is proposed to solve this problem, which is named domain transfer functional-link neural network (DT-FLNN). The framework of the proposed method includes two stages, and the issue of domain adaption is separately considered at each stage. In the first stage, the activation matrix of the source domain is reconstructed to narrow the distribution difference, and the augmented input matrices of the source and target domains are formulated. Then, the latent variable (LV) based linear regression method for transfer learning is performed at the second stage to train the FLNN of the target domain, and the task of domain adaption is realized by introducing a regularization term. Besides, a systematic method is also presented to determine the hyper-parameters in the proposed DT-FLNN method. The efficiency of the proposed method is evaluated by employing a numerical example and an industrial application. Compared with other nonlinear regression approaches for transfer learning, the proposed method can further increase the prediction accuracy and reduce the influence of noise.