Content determination by PSO-based LS-SVM regression

Near infrared (NIR) spectroscopy has rapidly developed into an important and extremely effective analysis method. With the use of spectroscopy, support vector machine (SVM) was used as regressor. It is well known that the selection of hyper-parameters including the regularization and kernel parameters is important to the performance of least squares support vector machine (LS-SVM). In this paper, the particle swarm optimization (PSO) is applied to select the LS-SVM hyper-parameters. Additionally, to construct the learning samples, a spectrum energy-based approach is proposed to determine the wavelength region where the observed data are used to train LS-SVM for the regression task. Concentration prediction of water-ethanol mixtures is used to verify the proposed methods. Experimental results show that LS-SVM with RBF kernel is superior to conventional methods including artificial neural network and partial least squares models.