目的 采用近红外光谱（near-infrared spectroscopy，NIRS）技术，结合机器学习算法，实现热毒宁注射液（Reduning Injection，RI）金银花和青蒿（金青）萃取过程中固形物含量（solid content，SC）的在线监测，并基于NIRS技术建立萃取终点判别模型，以提高金青萃取过程的质量控制水平。方法 采用NIRS技术，结合偏最小二乘法（partial least squares，PLS）和多元自适应回归样条（multivariate adaptive regression splines，MARS）算法，模型经过光谱预处理方法的优选及特征变量筛选，建立最佳SC的在线监测模型；采用支持向量机（support vector machine，SVM）算法建立异常光谱判别模型，通过移动块标准偏差法（moving block standard deviation，MBSD）算法建立萃取终点判别模型。结果 PLS和MARS模型性能优异，相较于PLS模型，MARS模型性能有所提升，预测相对误差（relative standard error of prediction，RSEP）由2.87%降低至2.64%，性能偏差比（ratio of performance to deviation，RPD）由15.953 0升至17.376 1，2种算法模型均具有模型性能好、预测精度高的优点；MBSD算法用于萃取终点的判断，可有效提升萃取效率。结论 NIRS技术结合PLS算法和MARS算法，均可用于RI金青萃取过程SC的在线监测，MARS模型性能更佳；采用MBSD方法进行萃取终点判断，方法简便易行，可以满足生产实际需求。

Objective To adopt near-infrared spectroscopy (NIRS) technology and machine learning algorithms, the on-line monitoring of solid content (SC) in the extraction of Jinyinhua (Lonicerae Japonicae Flos and Qinghao (Artemisiae Annuae Herba) (Jinqing) in Reduning Injection (RI) was realized, and the extraction endpoint discrimination model based on NIRS was established to improve the quality control level of the extraction process. Methods Using NIRS technology, combined with partial least squares (PLS) and multiple adaptive regression spline algorithm (MARS), the optimal model for online monitoring of solid content (SC) was developed through the selection of spectral pretreatment methods and feature variable screening. Support vector machine (SVM) algorithm was used to establish the abnormal spectrum discrimination model for data cleaning, and subsequently, the extraction endpoint discrimination model was established by moving block standard deviation (MBSD) method. Results Compared with the PLS model, the performance of the MARS model was improved, and the relative standard error of prediction (RSEP) was reduced from 2.87% to 2.64%, ratio of performance to deviation (RPD) increased from 15.953 0 to 17.376 1. Both of the two algorithm models have the advantages of good model performance and high prediction accuracy. The MBSD algorithm can be used to determine the extraction end point, which can effectively improve the extraction efficiency by about 18%. Conclusion NIRS technology, when integrated with both PLS and MARS algorithms, is effective for online monitoring of SC during the Jinqing extraction process of RI, with the MARS model showing superior performance. The MBSD method for endpoint determination is straightforward and meets the practical needs of production.

R283.6

国家科技部长三角科技创新共同体联合攻关项目(2023CSJGG1700);连云港市市重点研发计划(产业前瞻与关键核心技术)：基于PAT的中药浓缩和萃取过程反馈调控技术研究(CG2320)