白酒发酵阶段的多维数据融合与可解释机器学习智能识别

In the traditional solid-state fermentation of Baijiu, production control primarily relies on human expertise and retrospective offline analysis due to the lack of real-time, objective assessment methods. To address this limitation, this study introduces an intelligent identification method for fermentation stages, integrating multi-dimensional sensor data with explainable machine learning. The fermentation process was objectively segmented into three distinct stages using unsupervised learning. The validity of this segmentation was subsequently corroborated through multiple dimensions, including biochemical kinetics, Partial Least Squares Discriminant Analysis (PLS-DA), and flavor compound accumulation. Building on this, the performance of six machine learning algorithms was systematically evaluated. The Multilayer Perceptron (MLP) model demonstrated the best performance, achieving an identification accuracy of 99.48% on an independent test set. Furthermore, Shapley Additive exPlanations (SHAP) analysis provided model interpretability, revealing that dynamic features such as CO₂ concentration, pH, temperature, and humidity, along with their nonlinear interactions, were the primary drivers for the model's classification decisions. This scalable and cost-effective approach facilitates the transition from experience-based brewing to standardized, intelligent production, offering a robust technical solution for the Baijiu and broader traditional fermented food industries.