Feature Selection Using iPLS Combined with iNSGA-Ⅲ for Near-Infrared Spectroscopic Determination of the Acidity of Huangshui, a By-product of Chinese Baijiu Production

To address the inefficiency and complexity of traditional chemical methods for measuring the acidity of Huangshui (HS), this study proposed a rapid and non-destructive detection approach using near-infrared (NIR) spectroscopy combined with partial least squares regression (PLSR). The raw spectra were preprocessed by Savitzky-Golay convolution smoothing to reduce noise interference. To simplify the model and enhance the predictive performance, a hybrid feature selection strategy integrating spectral band selection and wavelength optimization was developed. First, interval partial least squares (iPLS), synergy interval partial least squares (SiPLS), and backward interval partial least squares (BiPLS) were employed to preliminarily identify characteristic bands related to acidity. Subsequently, a multi-objective optimization framework was introduced, incorporating an improved non-dominated sorting genetic algorithm Ⅲ (iNSGA-Ⅲ) with chaotic initialization and adaptive mutation operators for secondary wavelength refinement. Results demonstrated that the PLSR model based on 70 optimal wavelengths selected by iPLS combined with iNSGA-Ⅲ had the best predictive performance with higher coefficient of determination of prediction (R_p²) and lower root mean square error of prediction (RMSEP) of 0.9309 and 0.4394 mmol/100 g compared to 0.7576 and 0.8250 mmol/100 g for the full spectral model, respectively. This study provides a theoretical foundation for rapid, non-destructive monitoring of HS acidity during Baijiu fermentation.