Multivariate diagnosis of the variability of NIR spectrometers under industrial applications

Abstract

The transfer of NIR spectroscopy to industry relies on the possibility of real time identification of abnormal spectra as well as uncontrolled sources of variation. This study proposes an unsupervised procedure for the identification under an industrial application of daily events (general changes) and abnormal observations. It consists in defining a spectral database at the beginning of a season, performing a principal component (PC) analysis, and calculating the PC scores over time. Process control statistics (Hotelling T2, Q) are used for multivariate supervision of the industrial application. Within this procedure 10,400 average spectra of onion bulbs were evaluated identifying events in 12 out of 66 work dates, as well as spectral trends throughout the season of 2002.