Algoritmo das projeções sucessivas associado ao Kernel-PLS para calibração multivariada não linear

Abstract

In this work is reported, for the first time, the use of the Successive Projection Algorithm for interval selection (iSPA) combined to nonlinear data modeling by Kernel Partial Least Square (Kernel-PLS). This new approach, namely iSPA-Kernel-PLS, is a linkup between uninformative variable removed by SPA and noise reduction in nonlinear data by Kernel PLS. The performance of the proposed method was evaluated in three cases of study: (i) two simulated data to quantitation of the analyte in which concentration-analytical signal relation is quadratic and (ii) sugar and brix quantitation in sugar cane process control at different steps using near infrared spectroscopy (NIR) in transflectance mode. The nonlinear relationship between sugar/brix and NIR intensities was confirmed by appropriate statistical tests. When compared with full model (full spectrum), the proposed methods showed better results in terms of RMSE, REP and R2 for all case. In addition, significant bias is always absent in interval selection models based; this information is supported by analysis of elliptical joint confidence region. Therefore, the obtained results show that interval or variable selection, widespread in the linear calibration context, is a useful tool in nonlinear context too.