银杏叶含有丰富的活性物质，受到科研工作者的广泛关注，近年来银杏叶及其相关制品更是在世界医药产品中占有一席之地。我国作为银杏储量大国，却一直扮演着原料出口国的角色，无法进入高端市场。而质控评估手段相对滞后已经成为影响为我国银杏产业发展，阻碍其产品打入国际市场的一个瓶颈。

本工作利用核磁共振分析对样品适应性强、易于质控、提供信息丰富等特点，结合现代大数据处理理念和手段，提出了对银杏叶及其制品进行质量评价的新思路和方法。主要研究内容和成果如下：

1. 本工作根据核磁共振测量的技术特色和银杏叶及其提取物的样本特点，建立了银杏叶提取物核磁共振指纹谱的获取方法，并利用实际样品对该方法中的样本采集、样本提取、样品制备和¹H NMR谱获取等环节进行了方法学的考证。研究结果表明：建立的方法技术环节稳定可靠，所获取的¹H NMR谱满足了指纹谱的基本要求，可为银杏叶及提取物的评估提供科学依据。该方法实验流程简单易行，质控性强，具有很好的推广应用前景；不仅可以用于银杏叶及其提取物的评估，也为我国中草药及其产品的质地评估提供了新的思路。

2. 银杏叶提取物核磁指纹谱的谱学分析。结合多种核磁共振测试技术对所获取谱图中的信号进行了系统分析，确定了样品中的特征成分，完成了对指纹信号的确认。通过对信号强度的分区比对可以对银杏叶提取物样本间的差异性进行判别。分析结果表明：本工作所提出的方法获取的核磁谱提供了丰富的指纹信息，可以作为银杏叶及提取物样本评估的有效手段。

本工作还根据银杏叶提取物内化学成分复杂，¹H NMR指纹谱信号重叠严重的情况开发了HSQC相对定量分析方法。通过方法学检验证明了该方法在本工作应用的可行性，并完成了实际样品中萜内酯和黄酮类化合物等特征性化合物的相对含量的检测，进一步丰富了用以对银杏叶样本进行评估的数据信息。测试结果表明：HSQC定量不仅改善了数据的分辨率，且具有很好的稳定性，为复杂体系（如生物提取物）中某些特定成分的分析提供了一种有效的手段。

3. 建立了指纹谱的机器学习分析模式。利用主成分分析法和偏最小二乘法等多变量分析方法对银杏叶提取物的指纹谱数据进行类分判别分析，准确地实现了对样本的分类归组，同时完成了影响样本类分特征信号的表征，即完成了样品中左右归组类分特征成分的确认。类分结果显示：（I）景观型样本与道地产区专用样本可明显区别开来；（II）不同季节的样本亦可进行分组判别；（III）样本中特征组分（如黄酮类化合物、萜内酯类化合物）是左右类分的关键因素，通过后续HSQC定量法对表征的特征信号进行分析发现，各组样本间活性组分（萜内酯和黄酮苷类化合物）含量差异较大，最大可达到25%。说明原料的来源对银杏叶提取物产品的品质至关重要。

多变量分析的结果进一步证明了本工作所获取的¹H NMR谱指纹特征明显，完全可以作为样本评估的基础数据。而将机器学习分析用于指纹谱的分析不仅可以提高样品评估的效率，亦可避免了指纹谱分析过程中某些人为因素造成的误判，有利于保障样本评估的客观性。该工作表明核磁共振指纹谱结合机器学习分析是银杏叶及其提取物评估一种有效手段。

4. 建立银杏叶提取物分类识别的深度学习分析方法。不同于常规核磁共振数据的处理方法和分析模式，本工作首先将核磁共振的fid信号转化为彩色图片，而后利用深度学习方法对图片进行识别。本工作按照指纹谱的要求获取了176个银杏叶提取物样品（包括112个同一厂家不同批号的样品，和64个其它来源的样品）的¹H NMR数据，选用了目前国际上较为流行的三种深度学习分析模型（VGG16、Inception-v3、Resnet50）分别对这批数据进行分类识别尝试性分析。分析结果表明：经过优化后的三种模型均能用于对本实验样品的分类识别；其中以VGG16模型的处理效果最好，训练集和预测集准确率均可以达到95%以上，成功地实现了对银杏叶提取物样品的精确分类识别分析。深度学习分析充分展示了核磁共振检测指纹信息丰富的优势，弥补了目前银杏叶提取物产品检测中以单一化合物指标为评估依据带来的某些缺陷。该方法的引入和开发将为银杏叶产业及相关企业产品稳定性的监控及合格产品的识别提供一种新的思路和可行性方案。

本工作利用核磁共振技术的优势，深入研究了利用核磁共振技术对银杏叶及其制品进行质量控制的可行性，开发了核磁共振指纹谱的谱学分析、机器学习分析和深度学习分析的方法。尤其是深度学习分析方法的利用，开辟了核磁数据处理的新模式，亦是将现代大数据的理念和方法引入中草药评估体系中一个有益尝试。目前利用核磁共振技术进行复杂体系的质控研究在国内仍处于起步阶段，本工作的研究内容和研究成果将为银杏叶及其制品的质控工作提供新的工作思路和相关科学依据及技术保障。

Ginkgo biloba leaves are rich in active substances, so they are widely concerned by researchers. In recent years, Ginkgo biloba leaves and its related products occupy a place in the world's pharmaceutical products. China, as a large country of Ginkgo reserves, can only playing the role of raw material exporter, and can not enter the high-end market. However, the lag of quality control evaluation method has become a bottleneck for the development of ginkgo industry and hindering its products from entering the international market.

Based on the characteristics of NMR analysis, such as strong adaptability to samples, easy quality control and rich information, combined with the concept and means of modern big data processing, this work puts forward a new idea and method for quality evaluation of Ginkgo biloba leaves and its products. The main research contents and achievements are as follows:

1. According to the technical characteristics of NMR measurement and the sample characteristics of Ginkgo biloba and its extract, this work established the method of obtaining the NMR fingerprint of Ginkgo biloba extract, and used the actual samples to verify the methodology of the sample collection, sample extraction, sample preparation and ¹H NMR spectrum acquisition. The results show that the technology is stable and reliable. The ¹H NMR spectrum obtained by this method meets the basic requirements of fingerprint spectrum, which can provide scientific basis for the evaluation of Ginkgo biloba leaves and extracts. This method is simple and easy to use, has strong quality control, and has a good prospect of promotion and application. It can not only be used for the evaluation of Ginkgo biloba leaves and its extract, but also provide a new idea for the texture evaluation of Chinese herbal medicine and its products.

2. Spectral analysis of Ginkgo biloba extract. Combined with a variety of nuclear magnetic resonance testing technology, the signal in the obtained spectrum was analyzed systematically, the characteristic components in the sample were determined, and the fingerprint signal was confirmed. The difference between samples of Ginkgo biloba extract can be distinguished by the ratio of signal intensity. The results show that this work can provide abundant fingerprint information for the NMR spectrum obtained by the proposed method, and can be used as an effective method for the sample evaluation of Ginkgo biloba leaves and extracts.

In this work, the quantitative analysis method of HSQC was developed according to the complex chemical composition of Ginkgo biloba leaves and the serious overlapping of ¹H NMR fingerprint signals. The feasibility of this method was proved by the method test, and the relative content of terpene lactones and flavonoids in the actual samples was detected, which further enriched the data information for the evaluation of ginkgo leaf samples. The results show that HSQC quantitative analysis not only improves the resolution of data, but also has good stability. It provides an effective method for the analysis of some specific components in complex system (such as biological extract).

3. A machine learning analysis model of fingerprint spectrum is established. By using the PCA and PLS method and other multivariable analysis methods, the fingerprint data of Ginkgo biloba extract were classified and analyzed, and the classification and grouping of samples were realized accurately. At the same time, the characterization of the signal affecting the classification characteristics of samples was completed, that is, the confirmation of the characteristic components of the left and right classification components in samples was completed. The results show that: (I) landscape samples can be distinguished from the special samples in the real estate area; (II) samples in different seasons can also be grouped; (III) the characteristic components (such as terpene lactones) in the samples are the direct factors of the left and right components. Through the follow-up HSQC quantitative analysis of the characteristic signals, it is found that the active components (terpene lactones and yellow) among the samples of the components can be identified. The content of ketosides is quite different, up to 25%. It shows that the source of raw materials is very important to the quality of Ginkgo biloba extract.

The results of multivariate analysis further prove that the fingerprint features of ¹H NMR obtained in this work are obvious, which can be used as the data base of sample evaluation. The application of machine learning analysis in fingerprint spectrum analysis can not only improve the efficiency of sample evaluation, but also avoid the misjudgment caused by some human factors in the process of fingerprint spectrum analysis, which is conducive to the objectivity of sample evaluation. This work shows that NMR fingerprint combined with machine learning analysis is an effective method to evaluate Ginkgo biloba leaves and its extract.

4. Establish a deep learning analysis method of Ginkgo biloba extract classification and recognition. Different from the processing method and analysis mode of conventional NMR data, the deep learning analysis model first transforms the FID signal of NMR into color image, and then uses convolution neural network algorithm to recognize the image. According to the requirements of fingerprint spectrum, 176 samples of Ginkgo biloba extract (including 112 samples from the same factory with different batches, and 64 samples from other sources) were collected for ¹H NMR data, and three popular deep learning analysis models (VGG16, Inception-v3 and Resnet50) were selected to classify and identify the data. The results show that all three models can be used for the classification and recognition of the samples in this experiment, among which VGG16 model is the best, and the accuracy of training set and prediction set can reach more than 95%, which successfully realizes the accuracy classification and recognition analysis of Ginkgo biloba extract samples. Deep learning analysis fully demonstrated the advantages of NMR fingerprint information, and made up for some defects in the current Ginkgo biloba extract product detection based on a single compound index. The introduction and development of this method will provide a new idea and feasible scheme for the monitoring of product stability and the identification of qualified products in Ginkgo biloba leaves industry and related enterprises.