Active probabilistic sample selection for intelligent soft sensing of industrial processes

Abstract

This paper proposes a new active learning strategy based soft sensor upon the Gaussian process regression (GPR) model, in order to improve the prediction performance under a limited number of labeled data samples. The main objective of the new soft sensor is to opportunely label data samples in such a way as to maximize the soft sensing performance while minimizing the number of samples used, and thus to reduce the costs related to human efforts. By taking advantage of the GPR model, the information of prediction uncertainty is used to make a new probabilistic sample selection strategy, upon which the active learning GPR model is formulated for soft sensing. Detained analyses and comparative studies are carried out between the active learning strategy driven GPR model and random selection strategy driven GPR model through an industrial case study.