Improving Featured-based Soft Sensing through Feature Selection

Abstract

Driven by the expanding applications of spectroscopic technologies, many advancements have been reported for soft sensor modeling, which infers a sample’s properties from its spectroscopic readings. Because the number of wavelengths contained in a sample spectrum is usually much larger than the number of samples, “curse-of-dimensionality” is a common challenge that would affect the predictive power of the soft sensor. This challenge could be alleviated through variable selection. However, there is no guarantee that the truly relevant variables would be selected, and the selected variables are often (very) sensitive to the choice of training and validation data. To help address this challenge, we have developed a feature-based soft sensing approach by adapting the statistics pattern analysis (SPA) framework. In the SPA feature-based soft sensing, the features extracted from different segments of the complete spectrum were utilized to build the model. In this way, the information contained in the whole spectrum is used to build the model, while the number of the variables is significantly reduced. In this work, by integrating a variable selection approach we developed recently with SPA, we not only further improve the soft sensor’s prediction performance, but also identify the key underlying chemical information from spectroscopic data. The performance of the improved feature-based soft sensing approach, termed SPA-CEEVS, is demonstrated using two NIR datasets, and compared with several existing soft sensing approaches.