Application of terahertz dielectric constant spectroscopy for discrimination of oxidized coal and unoxidized coal by machine learning algorithms

Abstract Coal spontaneous combustion has always been a global catastrophe, which threatens valuable coal resources, human safety and health, and the environment. The key to preventing coal spontaneous combustion is early monitoring at the low-temperature oxidation stage of coal. The terahertz dielectric property measurement system and machine learning algorithms were used to test the feasibility of a non-contact real-time fast discrimination of oxidized coal from its unoxidized counterparts. Logistic regression analysis (LRA), Support vector machines (SVM), and random forest (RF) models were applied to classify coals based on oxidation degree. The results proved that differences of dielectric properties at terahertz band between unoxidized and oxidized coals did exist. The best classification result with an accuracy of 87.50% in the prediction set was achieved using the SVM method based on Gaussian kernel function combined with low-frequency (75-110 GHz) terahertz spectrum of the imaginary part of dielectric constant. The results prove that terahertz dielectric constant spectrum, combined with machine learning algorithms, would be a promising technique to monitor coal spontaneous combustion with high efficiency.