目的 建立一种基于近红外光谱（near infrared spectroscopy，NIRS）技术的金银花Lonicerae Japonicae Flos产地鉴别和酚酸类成分含量预测的分析方法。方法 采集不同产地不同批次的金银花的NIRS数据，采用高效液相色谱法测定其中的酚酸类成分含量，然后采用多种数据预处理方法提高光谱数据的信号特征，利用特征变量筛选方法提取与检测目标相关的特征波段。利用K近邻（K-nearest neighbor，KNN）和线性判别分析（linear discriminant analysis，LDA）建立产地鉴别模型，采用加权软投票法整合KNN和LDA模型的预测结果。利用支持向量机（support vector machine，SVM）和随机森林（random forest，RF）算法分别建立绿原酸、异绿原酸A和异绿原酸C成分的含量预测模型，并从中为每种成分选取合适的建模方法。结果 采用加权软投票法构建的产地鉴别模型的预测性能最佳，对预测集的预测准确率为95.7%、对数损失值为0.214、Kappa系数为0.907，对巨鹿县、封丘县和平邑县3个产地的识别准确率分别为98.8%、83.3%和94.4%。构建的SVM模型对预测集金银花样品中的绿原酸成分的预测决定系数（coefficient of determination，R²）为0.810 2，均方根误差（root mean square error，RMSE）为0.313 6%，相对预测偏差（relative percent deviation，RPD）为2.305 2，对异绿原酸A成分的预测R²为0.804 0，RMSE为0.135 4%，RPD为2.268 2，构建的RF模型对异绿原酸C成分的预测R²为0.271 0，RMSE为0.052 0%，RPD为1.176 3。结论 NIRS技术结合化学计量学方法可以有效区分不同产地的金银花，并可以较为准确的预测绿原酸和异绿原酸A成分的含量，为金银花质量的快速、无损评价提供方法依据。

Objective To establish an analytical method for the identification of Lonicerae Japonicae Flos (LJF) origin and the prediction of phenolic acid contents based on near infrared spectroscopy (NIRS) technology. Methods NIRS data of different batches of LJF from different origins were collected, and the content of phenolic acid components was determined by high performance liquid chromatography (HPLC), and then multiple data preprocessing methods were used to improve the signal characteristics of the spectral data, and feature bands related to the detection targets were extracted by using the feature variable screening methods. K-nearest neighbour (KNN) and linear discriminant analysis (LDA) were used to establish the origin identification models, and the prediction results of KNN and LDA models were integrated using the weighted soft voting method. Support vector machine (SVM) and random forest (RF) algorithms were used to establish the content prediction models for chlorogenic acid, isochlorogenic acid A and isochlorogenic acid C components, respectively, and select appropriate modeling methods for each component. Results The origin identification model constructed using the weighted soft voting method had the best prediction performance, with a prediction accuracy of 95.7% a Log loss value of 0.214, a Kappa coefficient of 0.907 on the prediction set, and the identification accuracies for samples from Julu, Fengqiu, and Pingyi counties were 98.8%, 83.3%, and 94.4%, respectively. The constructed SVM model had a coefficient of determination (R²) of 0.810 2, a root mean square error (RMSE) of 0.313 6%, and a relative percent deviation (RPD) of 2.305 2 for predicting chlorogenic acid content on the prediction set of LJF samples. For predicting isochlorogenic acid A content, the R² was 0.804 0, RMSE was 0.135 4%, and RPD was 2.268 2. Additionally, the constructed RF model for predicting isochlorogenic acid C content with an R² of 0.271 0, RMSE of 0.052 0%, and RPD of 1.176 3. Conclusion The NIRS technique combined with chemometrics can effectively distinguish LJF from different origins, and can accurately predict the content of chlorogenic acid and isochlorogenic acid A components, which provides a methodological basis for the rapid and non-destructive evaluation of the quality of LJF.

R286.2

现代中医药海河实验室科研项目（22HHZYSS00003）