Fast Characterization of Biomass and Waste by Infrared Spectra and Machine Learning Models

Heterogeneity is a most serious obstacle for treatment and utilization of biomass and waste (BW). This paper proposed a fast characterization method based on infrared spectroscopy and machine learning models, thus to roughly predict the elemental composition and heating value of BW. The fast characterization results could be used to sort different BW components by their suitable downstream utilization techniques. The infrared spectra based hybrid model contained a feature compression section to extract core information from raw infrared spectra, a classification section to distinguish inorganic dilution, and a regression section to generate the elemental composition and heating value results. By parameters optimization, the accuracy of this hybrid model reached 95.54%, 85.53%, 92.40%, and 92.49% for C content, H content, O content, and low heating value prediction, respectively. The robustness analysis was conducted by completely rearranging the training and test sets, and it further validated the hypothesis that the infrared spectra contains enough qualifying and quantifying information to characterize these properties of BW. Compared with previous literature, the C-H, C-O, and O-H correlations in BW were also well kept in the predicted results. This work is hoped to enhance upstream sorting system design for treatment and utilization of BW.