A Deep Learning Model for Real-time Detection and Prediction of Moisture Content in the White Tea Withering Process Based on Near-Infrared Spectroscopy

This study proposes a novel deep learning model, named STA-BiGRU-XGBoost, for predicting moisture content during the white tea withering process using near-infrared spectroscopy. The model integrates spatiotemporal attention mechanisms, bidirectional gated recurrent units (BiGRU), and the XGBoost algorithm to address challenges such as extended withering durations, environmental variability, and temporal variations in spectral data. The Maximum Relevance Minimum Redundancy (mRMR) algorithm is employed to select critical variables, including hot air velocity, air duct temperature, and spectral absorbance. A spatial attention mechanism enhances feature relevance, while BiGRU captures long-term temporal dependencies. Temporal attention further adjusts the weights of key time steps. XGBoost is incorporated to improve model robustness against noise. Experimental results using a production-line dataset demonstrate that the STA-BiGRU-XGBoost model achieves superior prediction performance, with RMSE = 0.0920, MAE = 0.0772, and R² = 0.9806, significantly outperforming traditional models. Furthermore, the model’s interpretability is validated through attention weight visualization, highlighting key features associated with moisture evaporation dynamics.