Coupling Kernel Principal Component Analysis with ANN for Improving Analysis Accuracy of Seven-component Alkane Gaseous Mixture

Abstract

To further improving the analysis accuracy of Artificial Neural Networks (ANN) model for quantitative analysis of seven-component alkane gaseous mixtures composed of methane, ethane, propane, isobutane, n-butane, isopentane, and n-pentane, the Kernel Principal Component Analysis (KPCA) technique was proposed to couple with it. The gaseous mixtures were measured by a novel Acousto-Optic Tunable Filter Near Infrared (AOTF-NIR) spectrometer. KPCA mapped the NIR spectral data of gaseous mixtures by a Gaussian kernel to a high-dimensional feature space and implemented feature extraction in it. As input variables, the extracted features were fed into a three-layered ANN to create quantitative analysis model of above-mentioned seven component gases. The performance of KPCA-NN model was assessed by Root Mean Square Error of Prediction (RMSEP) of testing set. The RMSEP of seven components by KPCA-ANN were less than 0.361%. Comparing with the ANN model without KPCA feature extraction, the KPCA-ANN model obtained the less RMSEP values. The research results indicated that the KPCA-NN model shows higher analysis accuracy than ANN model.