ENBIS: European Network for Business and Industrial Statistics

A robust calibration scheme for heterogeneous spectral data from the animal feed industry

22 September 2009, 16:15 – 16:35

Abstract

Submitted by

Sabine Verboven

Authors

Sabine Verboven(a), Peter Goos(a), Mia Hubert(b)

Affiliation

(a) University of Antwerp, Belgium; (b) Catholic University of Leuven, Belgium

Abstract

Statistical methods used in regression modeling are very sensitive to outlying observations. These outlying observations are typically due to malfunctioning measurement equipment, human errors in the data recording process, or undesirable changes in the production process. There exist alternative methods that are not sensitive to outlying observations, so-called robust regression methods, but they are seldom used in industry. In this contribution, we demonstrate the usefulness of robust calibration methods in the production of animal feed.

One of the main priorities in one of Belgium’s largest animal feed companies is predicting the humidity level from a large set of Near InfraRed (NIR) spectra, measured on different samples of animal feed. An accurate prediction of the humidity level would allow an effective quality control and a substantial reduction in the production cost. The high-dimensional nature of the data set necessitates the use of dimension reduction techniques like robust Principal Component Regression or Partial Least Squares Regression to build accurate prediction models.

We combine these methods with newly developed robust preprocessing techniques and achieve a decrease of about 15% in prediction error compared to classical Principal Components and Partial Least Squares regression. The robust approach offers the additional advantage that outlying samples are identified during the different steps of the analysis and thus provide a better insight in the heterogeneity of the data.

Return to programme