Neighborhood preserving regression embedding based data regression and its applications on soft sensor modeling

• NPRE is proposed to obtain the data regression relation among two datasets.
• With the local properties, a more faithful representation of the input data is given.
• An efficency nonlinear regression model is constructed by NPE.
• Considering the nonlinear relation between the process data, the kernel extension of NPRE is proposed.

In the present study, a new local-based data regression technique named neighborhood preserving regression embedding (NPRE) is developed and applied for soft sensor modelling. Unlike previous work on global modeling, the local-variation based neighborhood preserving embedding (NPE) provides stable and reliable description of the data characteristics. Taking such latent variables obtained by NPE as the input feature for data regression, NPRE is employed to construct soft sensor model and applied to industrial case to estimate some product qualities or key variables that are difficult to measure online. Besides, considering the nonlinear relation between the process data, the kernel extension of NPRE is also proposed. Two case studies on a fermentation process and a debutanizer column are provided to demonstrate the efficiencies of the proposed method in variable prediction. Based on the root mean square errors (RMSE) and correlation coefficient criterions, comparisons are also made with the global-based soft sensors. The results illustrate that the proposed NPRE can achieve significant improvement in terms of prediction accuracy and data correlation for the nonlinear processes.