Reducing Cost of Process Modeling through Multi-source Data Transfer Learning

The availability of an accurate model can facilitate tasks such as design or monitoring in chemical industries. A well-constructed model is invaluable but it can be costly to construct due to the expense in acquiring the labeled data. In the effort to reduce cost of building model, transfer of information using available sources has been considered. However, only the linear transfer has been considered for chemical processes in the reported literature. Moreover, the unlabeled data problem for transfer is not considered too. In this work, a new Gaussian process (GP) model based transfer learning approach for modeling is proposed. The aim is to leverage the statistical approach of the GP to transfer the knowledge from some available data of other processes for both labeled and unlabeled data. These data is used to provide information that is lacking to the process of interest. Case studies are presented to demonstrate the features of the proposed method.