Gaussian process model based multi-source labeled data transfer learning for reducing cost of modeling target chemical processes with unlabeled data

In chemical industries, many important tasks such as process design and monitoring rely on the availability of a good model. A high-performance data-driven prediction model is desired and requires labeled data. However, this can result in increased expenses in modeling because of the effort required to obtain the labeled data. The transfer learning (TL) approach has been considered to reduce the cost of acquiring labeled data but the case of unlabeled data in transfer learning for chemical process modeling has not been considered. A new Gaussian process (GP) model-based TL under the setting with unlabeled data is proposed in this work. By leveraging the predictive variance, the transfer of knowledge aims to increase the level of confidence in the prediction after transfer. The main contributions of the article include proposing a new transfer learning method under the setting with unlabeled data based on GP model, as well as the inclusion of threshold in the weighting of transfer. The use of GP model allows a statistical component to be taken into account in the transfer learning objective function whereas the threshold in the weighting of transfer acts as a mechanism that reject unwanted information is considered. The threshold thus provides a parameter in the consideration of the effectiveness of the transfer. The proposed method is demonstrated using a case study and its applicability to an industrial melt-index data is also shown.

Automated summary

This work proposes a transfer learning method based on Gaussian process model for scenarios without labeled data. By leveraging predictive variance, the transfer of knowledge aims to increase confidence in prediction. The method also introduces a threshold in the weighting of transfer to improve the effectiveness of transfer.