Full Wavelength Selection and Deep Spectra Feature Extraction Network for Near-Infrared Spectroscopy Quantitative Analysis

Near-infrared spectroscopy (NIRS) has been widely used in the chemical industry and has attracted much attention because of its fast and non-destructive characteristics. However, spectra are susceptible to interference from external environmental factors, leading to spectral overlap and noise problems. To solve these problems, spectral wavelength selection can be performed. Traditional selection methods require manually choosing the appropriate method based on spectral data and combining it with appropriate regression modeling, which is complicated and time-consuming. In contrast, this study proposes a novel spectral wavelength selection method that intelligently filters the entire wavelength based on the correlation between spectra and output variables. Simultaneously, it utilizes a deep spectra feature extraction module to obtain a more abstract spectral representation applied in regression prediction. Evaluation results on the training and test sets demonstrate the outstanding performance of this model: RMSE of 0.007796 and 0.01036 for the training and test sets, respectively, R2 of 0.9918 and 0.9842, with an impressive RPD of 7.9703. These metrics indicate that this model meets the rigorous requirements for high-performance models in the chemical engineering domain.