基于时频谱特征的白酒品质分类方法研究

为了建立快速、准确的白酒品质鉴别方法，利用机器学习方法对不同品质的白酒建模。为了提取不同品质白酒的特征，使用离子迁移谱对不同品质白酒进行分析，构建了基于白酒离子迁移谱信号的特征向量，并对不同品质的白酒进行了识别与分类。白酒样本的离子迁移谱信号通过利用美国Excellims公司GA2100型电喷雾-离子迁移谱仪（ESI-IMS）采集获得，每一个离子迁移谱信号是强度随时间变化的时间序列信号；提取了原始数据离子迁移谱的时域特征谱峰。为了获得更全面的特征，对离子迁移谱数据进行了傅里叶变换并提取频域内的特征谱峰。同时为了表述信号变化的特征，计算了离子迁移谱的谱熵和过零率，构建N×9维的特征向量矩阵；使用主成分分析(PCA)和线性判别分析(LDA)分别对上述获得的特征进行了特征降维，其中使用PCA对特征向量矩阵降维后的前三维特征对整体特征的累计贡献率达到了95%，而使用LDA对特征向量矩阵降维后的前两维特征对整体特征的累计贡献率就达到了95%。因此，选择了LDA作为特征降维方法；最后，利用机器学习中的非线性分类器支持向量机(SVM)对白酒离子迁移谱数据进行分类研究。实验结果表明，在真酒和添加酒精的白酒二分类中，SVM方法正确分类率达到100%；而在真酒和分别添加10%，20%，30%，40%和50%酒精浓度的五种假酒的六分类中，SVM方法正确分类率达到99.7%。比较了逻辑回归(LRM)分类、模糊C均值分类(FCM)和K近邻分类(KNN)对白酒样本离子迁移谱分类实验结果。研究表明，对于离子迁移谱非常接近的真酒和添加酒精的白酒，基于频谱特征向量的SVM方法能够准确的区分开来，为白酒的品质鉴别提供了一种新的检测方法。

Classification Method of Liquor Quality Based on Time and Frequency Spectrum Characteristics

In order to establish a fast and accurate method for liquor quality identification, this paper uses the machine learning method to model liquor of different quality. To extract the characteristics of different quality liquor, we analyzed different quality liquor with the ion mobility spectroscopy, constructed the feature vectors based on the signal of ion mobility spectroscopy, and classified different quality liquor. The ion mobility spectroscopy signals of liquor samples were obtained using the Excellims GA2100 electrospray ionization mobility spectrometry (ESI-IMS). Each ion mobility spectroscopy signal is a time series signal with its intensity varying with time. In the aspect of feature extraction, the time-domain characteristic peaks of the original data were extracted. Fourier transform is performed on the data of ion mobility spectroscopyfor more comprehensive characteristics, and the characteristic peaks in the frequency domain were extracted. At the same time, in order to express the characteristics of signal change, the spectral entropy and zero-crossing rate of ion mobility spectroscopy were calculated, and the N×9 dimensional feature matrix was constructed; Then, principal component analysis (PCA) and linear discriminant analysis (LDA) were used to reduce the dimensions of the features. The cumulative contribution rate of the first three-dimensional features of PCA to the overall features is 95%. By contrast, the cumulative contribution rate of the first two-dimensional features of LDA is 95%. Therefore, LDA is chosen as the feature dimension reduction method; Finally, a support vector machine (SVM), a nonlinear classifier in machine learning, was used to classify liquor ion mobility spectrum data. The experimental results show that the correct classification rate of the SVM method is 100% in the classification of real liquor and liquor with added alcohol; The correct classification rate of the SVM method is 99.7% in the six classifications of real liquor and five kinds of fake liquor with 10%, 20%, 30%, 40% and 50% alcohol concentration respectively. In addition, this paper compared the results of classification of ion mobility spectroscopy of liquor samples by logistic regression analysis (LRM), fuzzy C-means (FCM) and k-nearest neighbor (KNN). The results show that the SVM method based on spectrum feature vector can accurately distinguish the real liquor and the liquor with added alcohol, which provides a new detection method for identifying liquor quality.