OPTIMIZATION OF MULTIVARIATE CALIBRATION MODELS FOR THE PREDICTION OF OCTANE NUMBER IN MOTOR GASOLINE USING A CHEMOMETRICS EXPERT TOOLBOX AND SW-NIR

摘要

Near Infrared Analysis has been shown to be a viable technique for the measurement of multiple fuel parameters, including octane number in gasoline. In this paper, multiple chemometric techniques are applied to optimize octane prediction using third overtone near infrared (NIR) spectral data. Algorithms tested include linear principal component analysis (PCR) and partial least squares (PLS); non-linear forms of PCR and PLS; as well as locally weighted regression with PCR (LWR2), and locally weighted regression with PLS (LWR3). It has been suggested by some that prediction of octane number using NIR is a non-linear problem requiring non-linear methods. Weaknesses in current approaches involve the use of linear techniques for coping with non-linearities of the relationship between the physical parameter of octane number and the NIR spectral absorptions. The lack of fit due to non-linearities contribute more to the overall modelling error than the experimental error [reference engine error] of the calibration sample set. Using non-linear techniques for modelling the octane numbers from near-infrared spectra can improve the analytical performance and resistance of the model to noise in the instrument response (X-block) and reference values (Y-block) data. The non-linear techniques have resulted in 2-3 times improvements over PLS 1 (the generally accepted method of choice).