Semi-supervised Machine Learning Algorithm in Near Infrared Spectral Calibration: A Case Study to Determine Cetane Number and Total Aromatics of Diesel Fuels

Abstract

A new spectral calibration algorithm, Laplacian regularized least squares (LapRLS), was proposed. Commonly least squares support vector machine (LS-SVM) and partial least squares (PLS) are used for the spectral quantitative model establishment. However, LS-SVM and PLS are supervised machine learning algorithms which just make use of labeled data. LapRLS is a semi-supervised machine learning algorithm which makes use of both labeled and unlabeled data for training. In this study, LapRLS was used to establish the quantitative relationship between near infrared (NIR) spectra and cetane number (CN) and total aromatics of diesel fuels. Near infrared (NIR) spectroscopy is a widely used technique for monitoring chemical compounds in petroleum industry. A total of 381 obtained samples were randomly split into two sets under different proportion. One set was used as calibration set (labeled data) whereas the remaining samples were used as the prediction set (unlabeled data). LapRLS, LS-SVM, and PLS were used to establish determination models based on NIR spectra. Results show that the best performance of determination was achieved by LapRLS, which indicates that LapRLS can utilize unlabeled data effectively on NIR spectral data for the determination of the chemical compounds of diesel fuels.