Rapid prediction of cellulose, hemicellulose, and lignin content in biomass using FTIR based on transfer learning

The content of cellulose, hemicellulose and lignin varies significantly from one biomass to another, and this difference can seriously affect the chemical reaction activity of the biomass, so it is important to clarify the content of these three components in the biomass. Compared with traditional wet chemical method, Spectroscopy has been applied to the rapid detection of biomass components because of its advantages of simple sample handling, rapidity, and accuracy detection methods. In this study, FTIR combined with machine learning algorithms was used to achieve the rapid detection of biomass three-component content. Eighty biomass three-component content and FTIR spectral data were obtained and spectral preprocessed. Four machine learning models (PLSR, LS-SVR, RF, ELM) were constructed and analyzed for model performance comparison. The results showed that LS-SVR had the best performance, with an R 2 is 0.983 and an R MSE is 1.74 for predicting cellulose, an R 2 is 0.952 and an R MSE is 6.35 for predicting hemicellulose, and an R 2 is 0.949 and an R MSE is 2.67 for predicting lignin. However, when applied to new biomass datasets, the predictive model yielded an R 2 value below 0.663. To enhance the model's generalization capability, this paper proposes a transfer learning algorithm centered on ε-WLS-SVR. Using only 10 % transfer samples, the ε-WLS-SVR model achieved an R 2 exceeding 0.927 and an R MSE below 7.48, significantly improving its generalization performance. In this study, FTIR spectroscopy combined with ε-WLS-SVR algorithm was proposed to rapidly predict the cellulose, hemicellulose and lignin contents in biomass. It also investigates model transfer optimization, providing an accurate, fast, and green method for detecting biomass composition. • FTIR combined with machine learning accurately predicted the three-component content of biomass. • A transfer learning algorithm centered on ε-WLS-SVR enhanced the model's generalization capability. • T The SHAP algorithm was employed for interpretability analysis of the black-box model. • The R 2 values for predicting the three-component content using the FTIR combined with ε-WLS-SVR model all exceeded 0.927.