Transfer learning for end-product quality prediction of batch processes using domain-adaption joint-Y PLS

Abstract

In this work, a domain-adaption joint-Y partial least squares (JYPLS) is proposed to solve the problem of transfer learning for end-product quality prediction of batch processes. The difference from the variances in source and target domains is included as a regular term in the objective function of the traditional JYPLS model to realize the trade-off between minimizing the difference of empirical distribution in source and target domains and maximizing the covariance between latent variables and output variables. Since the issue of domain-adaption is considered in the proposed method, the prediction performances can be further improved. And the merits of JYPLS method can also be retained at the same time. The proposed method is tested on simulated data sets to verify the efficiency of the additional index in the objective function. It is also applied to predict the final mean particle size of a new cobalt oxalate synthesis process, and the data sets used to build the data-driven model is obtained from two synthesis processes with different model parameters and control policies. Compared with traditional JYPLS method, the prediction accuracy of the proposed method in the target domain has been greatly improved.