An accurate prediction of the physicochemical properties of traditional Chinese medicine granules using a multi-source data model fusion strategy based on deep ensemble learning algorithms

The quality control of Traditional Chinese Medicine Granules (TCMGs) is a critical aspect of industrial production. However, existing methods still face challenges in achieving rapid and accurate detection of their physicochemical properties, as well as ensuring efficient quality control. This study proposes a multi-source data model fusion strategy based on deep ensemble learning (DEL) to enhance the quality control of TCMGs. Forty batches of granule samples with varying physicochemical properties were prepared using Design of Experiments (DoE) and analyzed using conventional detection techniques, Fourier-transform near-infrared spectroscopy (FTNIR), and visible/near-infrared hyperspectral imaging (Vis/NIR-HSI), generating multidimensional datasets. Models were developed using chemometric methods, deep learning (DL), and DEL. The results demonstrate that DEL significantly improves the prediction accuracy and robustness of key indicators. For instance, in predicting the angle of repose and hesperidin content, DEL achieved RPD values of 4.174 and 4.829, and R2P values of 0.941 and 0.956, respectively. Compared to single DL models, DEL reduced the RMSEP/RMSEC ratios for the angle of repose, particle size, and hesperidin content to standard ranges. These findings highlight the substantial potential of DEL-based multi-source data model fusion strategies for analyzing complex datasets and achieving efficient quality control. This approach is extendable to other complex systems, enabling real-time monitoring and providing theoretical support for the intelligent and automated control of industrial production processes.