Final quality prediction method for new batch processes based on improved JYKPLS process transfer model

摘要

Data-driven methods have been successfully used in modern industrial production. The sufficient data is the basis for implementing these methods. However, it is often impossible to meet the requirement for a new industrial process. In this study, an improved JYKPLS (Joint-Y kernel partial least squares) process transfer model is proposed to solve this issue and perform final product quality prediction for a new batch process. Based on the latent variable transfer technology, the rich information from similar old process data is transferred to accelerate the proceeding of building a new process model. The requirements on the amount of modeling data and prior knowledge of new processes are visibly reduced. Moreover, in order to handle the nonlinear correlation in process data, the kernel function is introduced to make data linear and separable. With actual productions operating, the transfer model is improved gradually by updating it with online data. When the prediction error falls into its confidence interval, the old data with lower similarity will be eliminated to avoid the negative transfer. The prediction results of penicillin concentration verify the effectiveness of the proposed method.