Acta Scientiarum Polonorum Technologia Alimentaria

Background. Baijiu, a traditional Chinese distilled liquor, is produced using grains as the primary raw mate­rial. A key component in this process is jiupei (the fermented grain mash), which serves as the essential sub­strate, particularly in traditional solid-state fermentation methods, such as those employed in Baijiu production. Jiupei is a semi-solid mixture obtained by crushing steaming grains and mixing them with fermentation starters (qu), and then subjecting them to saccharification and fermentation in fermentation vessels, such as pits or jars.
Materials and methods. To address the need for real-time prediction of key physicochemical parameters – such as total acid, total ester, and total alcohol – during jiupei fermentation, a hybrid deep learning model was developed. This model integrates a Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM) network, and an attention mechanism. The concentrations of total acid, total ester, and total alcohol during fermentation were precisely quantified using a combination of chemical analysis and gas chroma­tography. Environmental variables, including temperature, humidity, carbon dioxide concentration, and pH, were then selected based on their significant associations with the target parameters as identified by Spear­man’s rank correlation coefficients. This non-parametric approach effectively addressed the challenges posed by non-normal data distributions during feature selection.
Results. Experimental results demonstrate that, compared to the single LSTM model, the proposed model improves the prediction determination coefficients (R²) for total acids, total esters, and total alcohols by 17.40%, 3.80%, and 5.70%, respectively. Additionally, the mean absolute error (MAE) and root mean square error (RMSE) remain consistently below 0.019 and 2.890, respectively, while the goodness of fit exceeds 0.73 across all parameters. These results confirm that the model significantly enhances predictive accuracy and robustness in modeling nonlinear fermentation processes.
Conclusion. Environmental factors, including temperature, humidity, carbon dioxide concentration, and pH, have a significant impact on key physicochemical indicators such as total acid, total ester, and total alcohol dur­ing the Baijiu fermentation process. These influences directly affect fermentation efficiency and the quality of base liquor. Consequently, monitoring these parameters offers real-time evidence to support the precise regula­tion of the fermentation process, with the potential to improve the consistency of base liquor quality and reduce batch-to-batch variation. Furthermore, this study provides a methodological framework for parameter predic­tion in other complex microbial fermentation systems, such as those used in Huangjiu and soy sauce production.