目的 基于决策树算法，深入挖掘热毒宁注射液金银花青蒿醇沉过程（金青醇沉）数据并探究潜在生产规律，提升该过程质量控制水平。方法 依托数字化中药提取工厂数据平台收集205批金银花和青蒿浸膏（金青浸膏）历史数据并整合成数据矩阵。将数据集随机划分为训练集和测试集后分别采用分类与回归树（classification and regression tree，CART）、随机森林（random forests，RF）和TreeNet算法建立金青醇沉过程模型，比较各模型性能并基于历史数据划分关键变量控制范围。结果 RF和TreeNet模型性能较好且性能接近，综合各模型分析结果得出醇提罐料液比及金银花浓缩收率为重要的影响因素，对重要变量进行依存度分析并优选批次，并以优选批次的金银花分配浸膏质量及加醇量做控制图，密度为1.11 g/cm³的金银花浸膏分配控制范围为557.92～639.62 kg，加醇量的控制范围为3.370～3.828 m³；密度为1.12 g/cm³的金银花浸膏的控制范围为540.4～616.9 kg，加醇量的控制范围为3.317～3.859 m³。结论 决策树算法建立的金青浸膏醇沉过程模型能够有效地挖掘潜在的生产过程规律，为生产过程的质量控制提升提供技术支撑。

Objective Based on the decision tree algorithm, the data of Jinyinhua (Lonicerae Japonicae Flos) and Qinghao (Artemisiae Annuae Herba) (LA) alcohol precipitation process of Reduning Injection (RI, 热毒宁注射液) were deeply mined and the potential production rules was explored to improve the quality control level. Methods Based on the digital extraction factory of traditional Chinese medicine, 205 batches of RI extraction batch records and LA alcohol precipitation process parameters records were collected and integrated into a data matrix. After the data set randomly divided into training set and test set, classification and regression tree (CART), random forest (RF), and TreeNet were used to establish LA alcohol precipitation process models respectively. The performance of each model was compared, and the control range of key variables was divided based on historical data. Results RF and treenet models had good performance and close performance. Based on the analysis results of each model, it was concluded that the material liquid ratio of alcohol extraction tank and the concentration yield of honeysuckle were important factors. The dependence of important variables were analyzed and the batches were optimized, and the quality and alcohol amount of honeysuckle distribution extract of the optimal batch was taken as the control chart. The distribution control range of honeysuckle extract with density of 1.11 was 557.92-639.62 kg, and the control range of alcohol dosage was 3.370-3.828 m³; The control range of honeysuckle extract with density of 1.12 was 540.4-616.9 kg, and the control range of alcohol dosage was 3.317-3.859 m³. Conclusion The application of decision tree algorithm in alcohol precipitation of LA extract can effectively explore the potential production process rules, and provide technical support for the improvement of quality control in the production process.

R283.6

国家新药创制科技重大专项（2013ZX09402203）