Level: Intermediate
David Autrique, Data Scientist, BASF
John Van Landschoot, Master Thesis Student, Ghent University
Process monitoring with in-line sensors is very common in chemical production plants. These measurements typically generate huge amounts of data. If utilized properly, this wealth of information leads to better process understanding and ultimately to production optimization. In past years, spectroscopy gained importance as a fast and reliable in-line chemical process monitoring technique to track product quality. Since spectral information typically consists of many strongly correlated variables, multivariate dimension reduction techniques are typically employed to explain the observed process variation. Two interesting though fundamentally different modeling candidates can be identified: Partial Least Squares (PLS) and a marquee feature in JMP 14, functional Principal Component Analysis (functional PCA). Whereas PLS would treat spectral data as collections of individual points, functional PCA describes a spectrum as a functional shape. A potential advantage of the latter, i.e., functional spectral treatment, stems from the combination of smoothing and dimension reduction. In contrast to PLS, functional PCA takes into account the order of the different wavelength intensities of a spectrum. In this talk we will use actual process data to demonstrate the modeling workflow for both techniques. Then we will compare the modeling performance of both techniques and share our insights with respect to practical usage.
