Dynamic soft sensors with active forward-update learning for selection of useful data from historical big database

Conventional static soft sensor is incapable of handling the dynamic of processes. With abundance of data, the problem of variable correlations and a large number of samples are encountered; moreover, the quality of the data for the construction of the soft sensors can be crucial for performance. An active learning strategy based on a latent variable model (LVM) to select representative data for efficient development of the dynamic soft sensor model is proposed. The uncertainty information for data selection is provided by the Gaussian process (GP) model. The developed LVM with the auxiliary GP model can handle the process dynamic. An active forward-update scheme which can update the soft sensor model in advance is proposed to reflect the current status of the process and improve the prediction performance without waiting for the quality measurements. Two case studies are done to demonstrate the features and the applicability of the proposed method.