Apport du bootstrap à la régression PLS : application à la prédiction de la qualité des gazoles

摘要

The objective of our modelling approach is a part of the quality control process of refining products. We build a prediction model of gas oil chemical properties from near infrared spectroscopy (NIR) of mid-distillates. The large number of explanatory variables and their high level of multicolinearity leads to the use of PLS (partial least squares) regression. Then, bootstrap techniques are used to determine prediction intervals. We consider the PLS model as a multiple linear regression on PLS orthogonal components to implement these bootstrap methods. Thus, we approximate the coefficients and prediction errors distributions.