红参中提取出的有效活性成分人参皂苷含量对后续产品的质量有重要的影响。 传统的红参提取质量控制化学检测方法成本高, 具有滞后性。 已有的研究表明快速无损的近红外检测方法用于红参提取过程具有可行性, 但现有方法依赖仪器自带数据处理软件, 无法满足生产实际的精度和速度需求。 为实现红参提取过程的快速、 精确监测, 提出将多种智能光谱筛选算法应用在近红外光谱建模中, 并对比不同光谱筛选算法的性能和稳健性。 以红参提取液中含量高的人参皂苷Rg1和含量较低的人参皂苷Rc为目标, 采集了三个不同批次前两次红参提取液样本128份, 在线获取1 000~2 499 nm波段近红外原始光谱吸光度数据, 并同时采用国标方法高效液相色谱法测定目标人参皂苷含量, 首先采用竞争适应性重加权采样法(CARS)、 无信息变量消除法(UVE)、 随机蛙跳算法(RF)和连续投影算法(SPA)四种波长筛选算法进行波长降维处理, 然后使用筛选后的波长建立线性偏最小二乘(PLS)定量模型, 并通过模型的均方根误差(RMSE)、 决定系数( R²)和预测相对分析误差(RPD)等来评估模型的性能。 从四种波长优选算法PLS建模结果可知, 经RF优选后, 建模特征波长变量下降为原来的0.67%, 红参提取液中人参皂苷Rg1和Rc含量的 R²都达到了0.94以上, 预测均方误差分别为0.024 6和0.013 5, 预测集相对分析误差达到了4.84以上, 降低了建模的难度, 提高了建模的精度; 将RF和CARS在原始光谱、 全光谱、 SNV预处理后的全光谱上建模对比, RF波长筛选算法建模模型的性能整体较好, 不同的光谱范围和预处理方法下性能影响较小, 稳健性好。 综上表明RF是红参提取液建模相对理想的波长筛选算法, 基于RF的PLS算法实现了对红参两次提取液的一次建模, 可用于提取液中人参皂苷成分含量的快速检测, 为药物的在线提取控制提供理论支撑。

As an effective active component in red ginseng extraction, ginsenoside content has an important impact on the quality of follow-up products. Traditional chemical detection quality control methods have high costs and time-delay. Existing studies have shown that the fast and non-destructive near-infrared detection method is feasible for red ginseng extraction. However, the existing methods heavily rely on the data processing software algorithm of the instrument, which cannot meet the actual production accuracy and speed requirements. In order to monitor the extraction process rapidly and accurately, a variety of intelligent spectral selection algorithms are applied in the near-infrared spectral(NIRS) modeling, and the performance and robustness of different spectral selection algorithms are compared in this study. In order to detect the high content of ginsenoside Rg1 and the low content Rc in the red ginseng extract, 128 samples of red ginseng extract in the first two times extracted were collected from three batches, 1 000~2 499 nm band NIRS data were obtained online, and the content of ginsenoside was determined by using the international standard high-performance liquid chromatography (HPLC). Firstly, the dimension of the input wavelength was reduced by using four wavelength selection algorithms, namely, competitive adaptive reweighting sampling (CARS), the uninformative variable elimination (UVE), random frog (RF) and successive projection algorithm (SPA). Then the selected wavelength was used for the linear partial least squares (PLS) quantitative model establishment. At last, the performance of the model was evaluated by the root mean square error (RMSE), coefficient of determination ( R²) and relative analysis error (RPD), etc. According to the PLS modeling results of four wavelength optimization algorithms, after RF optimization, the characteristic wavelength variable of the modeling decreased to 0.67% of the original, R² of the ginsenoside Rg1 and Rc content in red ginseng extract reached above 0.94, the RMSE of the prediction was 0.024 6 and 0.013 5 respectively, and the RPD of prediction set reached above 4.84, which reduced the difficulty of the modeling and improved the accuracy of modeling. From the comparison of RF and CARS modeling in the original spectrum, full-spectrum and SNV pretreated full spectrum, the overall performance of the RF wavelength selection algorithm model is better. Different spectral ranges and pretreatment methods have little impact on the performance and good robustness. In conclusion, RF is a relatively ideal wavelength selection algorithm for the modeling of red ginseng extract. PLS based on RF realizes the one-time modeling of two red ginseng extracts, which can be used to rapidly detect ginsenoside content in the extract. The study provides theoretical support for the online extraction control of medicine.