A novel method to realize multicomponent infrared spectroscopy gas logging based on PSO-split peak fitting-SVM

Abstract

Listen

Translate article

The recent introduction of infrared spectroscopy gas logging technology is of great significance for the timely discovery of oil and gas, the prediction of well kicks, the prevention of blowout and gas invasion, and the comprehensive evaluation of reservoirs. However, the accuracy and stability of the current quantitative analysis methods for gas logging measurement data are limited by the severe overlap of infrared absorption spectra and the difficulty of sample preparation. In this paper, a novel hybrid method based on particle swarm optimization split peak fitting and a support vector machine (PSS) is presented to address this issue. The method uses a single peak that satisfies a specific absorption line to fit the mixed spectrum and classifies and identifies the parameters of a fitted single peak to complete the “mathematical separation” of the overlapping spectrum. The IR spectral data of 761 samples for methane, ethane, propane, n-butane, iso-butene, n-pentane, iso-pentane and carbon dioxide were collected for quantitative analysis experiments. Repeated comparison experiments of different feature extraction methods under the same partial least squares (PLS) regression method show that the PSS model established with fewer samples still has better analytical accuracy and stronger robustness. At the same time, PSS can significantly improve the model accuracy under different regression methods to a similar level, reducing the dependence on regression methods. PSS combined with PLS can simultaneously measure the contents of eight components in a multicomponent mixture, and the predicted average root mean square error is 0.7242% for 60 mixed experimental gas samples and 0.2534% for 3 industrial standard gas samples. The proposed method has a high practical application value for popularizing IR spectroscopy gas logging technology and provides a new concept for the quantitative detection of substances with high molecular similarity.