Research on near-infrared spectral data analysis of liquor fermented grains based on a multilayer perceptron model

This study developed a rapid method for quantifying acidity, starch, and moisture in liquor-fermented grains using near-infrared (NIR) spectroscopy and machine learning. Spectral data underwent preprocessing (SNV, SG, 1stD) to correct baseline shift and noise. Competitive Adaptive Reweighted Sampling (CARS) selected characteristic wavelengths, reducing data dimensionality. A Randomized Grid Search (RGS) strategy efficiently optimized Multilayer Perceptron (MLP) model parameters, cutting the optimization time by 80.2%. The optimized MLP models achieved excellent predictive performance, with determination coefficients (R 2 ) of 0.9623, 0.9315, and 0.9054 for acidity, starch, and moisture, respectively, significantly outperforming traditional PLSR models. This method enables rapid, accurate measurement of key components, providing strong support for online monitoring and quality control in liquor production.