Comparison of data summarization and feature selection techniques for in-process spectral data

In this work, approaches to data summarization and feature selection are assessed for predicting the mechanical properties of a polymer product based on complex heterogeneous in-process data. Pressure and temperature data as well as Near Infrared (NIR) spectroscopy data were captured at different sampling frequencies during the process and used to predict the yield strength of the product. Direct interpretation of NIR spectra is recognized as an intractable problem in material processing and chemometric approaches are applied to build models which must be calibrated against lab-characterized response data. The low sampling rate of such lab characterization relative to in-process data capture raises the question of how best to summarize the process data when predicting the material properties. Further, conventional Principal Component Regression (PCR) and Partial Least Squares (PLS) regression chemometric methods lack interpretability of the model and do not provide much insight for how best to control the process. In this work we compare two different approaches to data summarization and compare two different Recursive Feature Elimination (RFE) methods for feature selection. It is shown that RFE using Random Forest regression with data summarized over the entire production run yields the best predictive performance. It also delivers a sparse model in the original features which facilitates interpretation of physio-chemical changes in the material and provides useful insight for process control.