Abstract:As a critical step in the manufacturing of herbal medicines, extraction process is mainly used to collect chemical components and has a significant impact on downstream operations. However, the monitoring of extraction process is still lack of appropriate strategies and methods. As for herbal medicines with multiple components, the extraction of which tends to be a "black box" operation. In this work, the method based on near-infrared spectroscopy combined with machine learning was proposed to determine the extraction end-point and characterize the real extraction status of Gegen Qinlian Decoction. Partial least squares (PLS) regression, extreme learning machine (ELM), and Gaussian process (GP) regression were used to develop the calibration model. Additionally, a calibration-free method, the moving block standard deviation algorithm (MBSD) was also applied for end-point determination. Compared to the PLS regression and ELM models, the GP regression models could realize the accurate prediction of the extraction end-points of puerarin, berberine hydrochloride, baicalin, glycyrrhizic acid and soluble solids. The result given by the MBSD method further confirmed the reliability of the GP regression models. To address the issue of real-time monitoring of extraction process, the multivariate statistical process control (MSPC) model combined with three statistics (principal component score, Hotelling T 2 and distance to model X) was developed and used for the malfunction identification during the extraction process. The results showed that the five types of malfunctions were all detected using the MSPC model. In summary, the methods established in this paper may serve as a potential strategy for the real-time monitoring of the extraction procedure in complex chemical systems.

NIR spectroscopy as a PAT tool for the extraction process of Gegen Qinlian Decoction: End-point determination and fault detection