An Interpretable Prediction Method for Tobacco Drying Process Based on CGTNN

Abstract

The moisture content of tobacco is a critical quality indicator in the tobacco processing, it is essential to control the moisture content of the tobacco within an appropriate range through the drying process. However, production often faces large delays in moisture measurement, leading to the difficulty of anomaly detection and moisture control. This paper proposes a novel approach combining Correlation Graph and multi-scale Temporal Neural Networks(CGTNN) for real-time moisture prediction. The model transforms the raw data into graph data based on correlation coefficients, extracting the features in the production data from both spatial and temporal perspectives according to the strength of the relationships between different production processes. In addition, to address the risks posed by the uninterpretability of deep learning models,shapley values are used for interpretability analysis, aligning predictions with production experience. Experiments with production data show that the model accurately predicts moisture content with a Mean Absolute Error (MAE) of 0.016%, Root Mean Squared Error (RMSE) of 0.024%, and an Explainable variance (R2) of 0.987, outperforming other models. This approach significantly reduces delays and errors in moisture monitoring, enhancing accuracy and reliability in practical applications.

Keywords

• Tobacco drying
• graph neural networks
• correlation analysis
• spatial and temporal analysis
• deep learning interpretability