不同产地丹参药材的质量差异显著, 亟需建立准确、 快速的分析鉴别方法, 对丹参药材的产地进行识别。 激光诱导击穿光谱（LIBS）具有快速、 实时、 高效、 等特点, 克服了传统分析方法时间长、 程序复杂等问题。 人工神经网络（ANN）方法则具有强大的学习和泛化能力, 是一种快速、 准确的分析方法。 采用LIBS技术结合ANN方法构建了对不同产地丹参药材的鉴别方法。 实验首先收集来自安徽、 甘肃等六个不同产地的丹参样品, 并通过LIBS光谱仪对丹参样品进行光谱采集； 之后对丹参LIBS光谱的元素特征峰进行比对, 发现不同产地丹参样品的元素发射谱线强度存在着差异, 如Fe元素（238.20, 373.71 nm）和Ca元素（315.89, 317.93 nm）等； 采用最大最小归一化（MMN）、 标准正态变换（SNV）、 均值中心化（MC）、 Savitzky-Golay平滑滤波（SG）以及多元散射校正（MSC）五种预处理方法对LIBS光谱数据进行预处理优化, 减少光谱噪声以及其他干扰信息的影响； 最后分别搭建ANN分类模型, 从测试集分类准确率、 每类产地的敏感性、 精确率和特异性等方面进行比较, 选择最优模型。 基于原谱的ANN模型测试集分类准确率为94.24%； SNV、 MC两种方法并没有提升ANN模型的分类能力； MMN、 SG及MSC三种预处理方法均提升了ANN的分类效果。 SG-ANN模型取得了最佳鉴别效果, 外部测试集分类准确率为98.15%, 同时具有更高的敏感性、 精确率和特异性, 其中, 安徽、 河南两地丹参样品的判别结果最好, 敏感性、 精确率及特异性均达到100.00%, 其余四种产地丹参样品的敏感性、 精确率及特异性也在95.00%以上。 该结果表明, 选择合适的光谱预处理方法, 能显著提升ANN模型对于丹参产地的预测分类能力, 构建一种相关性更强的定性分析模型。 研究结果表明LIBS技术结合人工神经网络方法是一种很有前景的丹参药材分析鉴别方法, 为中药材质量监督体系提供一种新思路。

The quality of Salvia miltiorrhiza in different origins varies greatly, and it is urgent to establish an accurate and rapid analytical method for discrimination. Laser-induced breakdown spectroscopy (LIBS) has the advantages of fast, real-time, high efficiency, which overcomes many problems of traditional analysis methods. Artificial neural network (ANN) has strong learning and generalization abilities, a fast and accurate analysis method. Therefore, a method for discriminating Salvia miltiorrhiza from different geographical origins was developed by using LIBS combined with ANN. In the experiment, the samples of Salvia miltiorrhiza from six different origins, such as Anhui and Gansu provinces were collected, and the spectra of Salvia miltiorrhiza samples were collected by LIBS spectrometer. Then, comparing the element characteristic peaks of LIBS, it was found that there are differences in the element emission intensity of Salvia miltiorrhiza from different origins, such as Fe (238.20, 373.71 nm) and Ca (315.89, 317.93 nm). A supervised classification model was established by the ANN method combined with 5 different spectral preprocessing methods: maximum and minimum normalization (MMN), mean centralization (MC), standard normal transformation (SNV), Savitzky-Golay smooth filtering (SG) and multiple scattering correction (MSC). The RAW-ANN model has achieved a test set classification accuracy of 94.54%; SNV and MC methods did not improve the classification ability of the ANN model; And the three preprocessing methods of MMN, SG, and MSC all have improved the classification performance of the ANN model. The SG-ANN model achieved the best classification effect, with a test set classification accuracy of 98.15%. At the same time, it has higher sensitivity, precision and specificity, of which Anhui and Henan provinces have the best discrimination results, with sensitivity, precision and specificity reaching 100.00%. The other four orgins sensitivity, precision and specificity are also above 95.00%. The results showed that selecting an appropriate spectral preprocessing method could significantly improve the classification ability of the ANN method and build a more relevant qualitative analysis model. The above results show that LIBS combined with ANN is a promising method for analysing and identifying Salvia miltiorrhiza, which provides a new idea for the quality supervision system of Chinese medicinal materials.