An auto encoder as a feature-extraction tool for Raman spectroscopic compositional analysis of hydrocarbon mixtures and investigation of correlations of auto encoder–extracted variables with component concentrations

Abstract

This paper examines the ability of an auto encoder (AE) to serve as a feature-extraction tool in Raman spectroscopic compositional analysis of mixtures of 11 hydrocarbons (n-hexane, n-heptane, n-octane, n-nonane, isooctane, cyclohexane, methyl cyclohexane, benzene, toluene, xylene, and indan) and explores its potential application to vibrational spectroscopic analysis of petroleum and petrochemicals. We found the use of AE-extracted variables enhanced analytical accuracy when determining the concentrations of the four linear hydrocarbons and xylene compared with those using the original spectra. Although the peaks of the linear hydrocarbons were relatively indistinct and similar to each other, the AE effectively extracted the relevant features from the highly overlapped spectra, improving the accuracy. To investigate how the AE variables describe variation in the concentrations of the components, correlations between the sets of AE variables and component concentrations were examined using canonical-correlation analysis. The resulting absolute R values were high, ranging from 0.936 to 0.979. This suggests that the AE variables are mutually complementary and can track spectral variation induced by changes in the concentration of the components.