FTNIR光谱结合化学计量学用于白酒品牌鉴别的快速判别分析

发布时间：2018-04-10 15:29

本文选题：白酒品牌鉴别 + 傅里叶变换近红外光谱　；参考：《暨南大学》2016年硕士论文

【摘要】：中国是白酒生产、消费大国,知名白酒品牌鉴别是我国酒类产品质量检测的一个重要项目。传统的分析手段需要化学试剂、方法复杂、专业性强,不适合于大规模的应用。本文基于傅里叶变换近红外(FTNIR)光谱,将移动窗口波段选择分别结合到主成分线性判别分析(PCA-LDA)和偏最小二乘判别分析(PLS-DA),提出了两种用于光谱模式识别的集成优化方法,即MW-PCA-LDA和MW-PLS-DA,成功应用于我国知名白酒品牌泸州老窖的快速判别分析。收集到非泸州老窖(52%vol,浓香型)样品200个(阳性)和用于对照的泸州老窖(52%vol,浓香型)样品160个(阴性),分别利用1mm、2mm、5mm、10mm光程的样品池测定样品的FTNIR光谱。为了建立稳定可靠的模型,将全体样品随机划分为建模、检验集,建模集再随机划分为定标、预测集共30次。基于30次不同的划分的最优预测准确率(P_RECAve)优选模型参数。首先,在全谱波段(15000-4000 cm-1)上建立模型。采用PCA-LDA方法,最优P_RECAve为98.1%,最佳光程为2mm和10mm;采用PLS-DA方法,最优P_RECAve为100.0%,最佳光程为2mm。但是,采用的波数个数(N)高达2852,因此,进一步提取有效波数、降低模型复杂性尤为必要。其次,基于最佳光程(2mm、10mm)的光谱数据进行波段优化。采用MW-PCA-LDA方法,最优模型的光程为2mm,波段为5235-5130 cm-1,N为28,P_RECAve为100.0%;采用MW-PLS-DA方法,最优模型的光程为2mm和10mm,波段为5238-5204 cm-1和7186-7128 cm-1,N为10和16,P_RECAve均为100.0%。最后,采用不参与建模的检验样品,对三个最优模型进行检验,最优MW-PCA-LDA和MW-PLS-DA(2mm光程)模型的阴性、阳性检验识别率V_REC—、V_REC+均为100.0%,最优MW-PLS-DA模型(10mm光程)的V_REC—、V_REC+分别为98.3%和100.0%。结果表明:FTNIR光谱结合MW-PCA-LDA和MW-PLS-DA方法可用于白酒品牌鉴别的高精度判别分析。所提出的波段选择方法可以提取有效信息波长、降低模型复杂性,为设计小型专用光谱仪器提供有价值的参考。本文所提出的分析方法简单、快速、有效,在我国酒类产品质量检测具有重要应用前景。
[Abstract]:Chinese is liquor production, consumption, famous liquor brand identification is an important project for quality inspection of China's liquor products. The analysis of the traditional means of need of chemical reagents, method of complex, professional and strong, not suitable for large-scale applications. In this paper, Fourier transform infrared spectroscopy (FTNIR), based on the moving window band selection respectively according to the principal component linear discriminant analysis (PCA-LDA) and partial least squares discriminant analysis (PLS-DA), two for integrated optimization method of spectral pattern recognition is proposed, namely MW-PCA-LDA and MW-PLS-DA, fast discriminant analysis is successfully applied to China's well-known liquor brands. Luzhou Lao Jiao collected non Luzhou Lao Jiao (52%vol, Luzhou) sample 200 a (positive) and used to control the Luzhou Lao Jiao (52%vol, Luzhou) 160 samples (negative), respectively by 1mm, 2mm, 5mm, 10mm of the sample pool path determination of FTNIR spectra of the samples in order to. To establish a stable and reliable model, all samples were randomly divided into modeling, test set, and then randomly divided into modeling set calibration and prediction set a total of 30 times. 30 different classification accuracy based on optimal prediction (P_RECAve) optimize the parameters of the model. First of all, in the whole band (15000-4000 cm-1) on the model. Using PCA-LDA method, the optimal P_RECAve is 98.1%, the optimal path for 2mm and 10mm; using PLS-DA method, the optimal P_RECAve is 100%, the best path is 2mm. but the number of wave number a (N) up to 2852, therefore, further extraction of the effective wave number, it is necessary to reduce the model complexity. Secondly, based on the best path (2mm 10mm) band optimization of spectral data. By using MW-PCA-LDA method, the optimal path model for the 2mm band, 5235-5130 cm-1, N 28, P_RECAve 100%; using MW-PLS-DA method, the optimal path model for the 2mm and 10mm band, 5238-5204 C M-1 and 7186-7128 cm-1, 10 N and 16 P_RECAve were 100.0%., finally, the model does not participate in the test sample, to test three optimal model, optimal MW-PCA-LDA and MW-PLS-DA (2mm path) model test negative, positive recognition rate V_REC, V_REC+ was 100%, the optimal MW-PLS-DA model (10mm path) the V_REC and V_REC+ were 98.3% and 100.0%. respectively. Results show that it can be used in high precision liquor brand identification discriminant analysis FTNIR spectroscopy combined with MW-PCA-LDA and MW-PLS-DA method. The proposed method can extract the effective information of band selection wave length, reducing model complexity, and provide valuable reference for the design of small special instruments. The spectral analysis method this simple, fast, effective, has important application prospect in the quality inspection of China's liquor products.

【学位授予单位】：暨南大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：O657.33;TS262.3

【相似文献】