【作者】 冯利辉；

【导师】 张国文； 刘波平；

【作者基本信息】 南昌大学 ， 食品科学， 2010， 硕士

【摘要】 近红外光谱技术是近年来发展起来的一种快速、无损、方便的检测技术,其应用越来越广泛。化学计量学是一门新兴的交叉学科,在提取分析信息、光谱预处理以及模型的建立等方面具有自身的优势。近红外光谱技术与化学计量学的结合在定性判别和定量分析检测中都有很重要的应用,为食品安全检测提供了解决问题的新途径和新方法。本文共分为七章,主要研究了化学计量学结合近红外光谱技术用于食用植物油掺伪的定性鉴别和定量检测,以及食用植物油种类鉴别,酸价、过氧化值等指标的检测。1.综述了食用植物油种类判别、酸价、过氧化值以及食用植物油掺伪检测的国内外研究进展。概括介绍了近红外光谱技术原理以及近红外光谱技术在定性和定量检测中的研究现状。2.对大豆油、玉米油、花生油、芝麻油、山茶油等5种食用植物油种类鉴别的近红外光谱法进行了研究,分别采用马氏距离聚类分析方法和自组织竞争神经网络建立了判别模型,并探讨了光谱波长范围和光谱预处理方法对模型的影响,利用所建立的马氏距离聚类分析模型和自组织竞争网络模型对预测集25个样本进行预测,两个模型的预测准确率都达到100%。3.研究了偏最小二乘法(PLS)法和PLS-BP网络法结合近红外光谱技术用于同时检测食用植物油酸价和过氧化值,分别建立了酸价和过氧化值的定量分析模型。并分别采用所建立的酸价和过氧化值PLS模型和BP网络模型对预测集样本进行预测,结果表明,PLS所建模型对两组分预测集样品预测决定系数R2分别为0.9837和0.9752,预测均方根误差(RMSEP)为0.0752和0.00972,BP神经网络模型对预测集样品的R2分别为0.9695和0.9744,RMSEP分别为0.0595、0.00991。两种方法基本可以满足酸价和过氧化值同时测定的需要。4.研究了大豆油、玉米油、葵花籽油掺入到山茶油中的掺伪二元体系定性鉴别和定量测定。采用马氏距离聚类分析方法分别建立了山茶油掺伪与否的判别分析模型和三种掺伪山茶油分类的判别模型。两种掺伪判别模型的判别准确率都在99.1%以上,模型的预测效果满意。并建立了掺伪山茶油二元体系中大豆油、玉米油、葵花籽油含量的PLS模型,PLS模型的校正相关系数分别为0.99957、0.99962、0.99975；RMSEC分别为0.300、0.309和0.255。对预测集样本预测的RMSEP分别为0.467、0.272、0.410,同时预测集的真实值和预测值配对t-检验结果显示,差异均不显著,所得结果满意。5.配制掺有大豆油、菜籽油、棕榈油的掺伪花生油二元体系,采用PLS-BP网络法研究了花生油掺伪的定性鉴别和定量测定。采用BP网络建立了花生油掺伪鉴别模型,BP网络经744步左右达到训练目标,所建模型对校正集和预测集样本判别准确率分别为100%和96.0%。同时,采用BP网络建立起掺伪花生油中大豆油含量、菜籽油含量和棕榈油含量的定量校正模型,并对预测集样本进行结果预测,R2分别为0.9851、0.9901、0.9850,RMSEP分别为1.05、1.10、1.72。同时与PLS法和PCR法模型结果进行了比较,结果显示,PLS法预测结果稍好于BP网络模型,两种方法都可以满足掺伪花生油检测的需要。6.采用自组织特征映射(SOM)网络建立了掺有大豆油、菜籽油、花生油的掺伪芝麻油二元体系判别分析模型。采用PCA法对光谱数据进行压缩提取主成分,所建SOM网络训练500后判别准确率达到100%,利用SOM网络模型对预测集判别准确率为95.6%。此外,还采用PLS法对掺伪油含量进行定量分析并建立了定量校正模型。利用所建立的PLS模型对预测集样品的大豆油、菜籽油和花生油含量进行预测,预测结果的RMSEP分别为1.05、1.31和0.686,R2分别0.9951、0.9825和0.9944,预测结果满意。另外,本文还对芝麻油掺伪的三元体系(掺有大豆油和花生油)定量检测进行了探讨,建立了大豆油、花生油含量的PLS法模型,大豆油和花生油含量的PLS定量模型的校正相关系数分别为0.95714、0.97025,RMSEC分别为1.54和1.42。相关性还不够理想,定量分析的精确度尚须提高。更多还原

【Abstract】 Near-infrared spectroscopy (NIRS) is a fast, non-destructive and convenient analytical technique which applied in more and more fields. Chemometrics is a rising Cross-disciplinary with many advantages in the aspects of extracting, analyzing information, spectra pretreatments and building the models. To afford new methods to solve the problem for food security and detection, near infrared spectroscopy combined with chemometrics are very important applications for the qualitative and quantitative analysis.. This paper is divided into seven chapters, with the major study on qualitative discriminate for adulteration of edible vegetable oil and categorys of edible vegetable oil, quantitative detection for adulteration of edible vegetable oil, acid value and peroxide value by chemometrics and near infrared spectroscopy technology.1. The research on category identification, acid value, peroxide value and adulteration detection of edible vegetable oil were summarized at home and abroad. The principles of near infrared spectroscopy were introduced. At the same time the research status of the qualitative and quantitative detection by using near infrared spectroscopy technology was concluded.2. The category identification of edible vegetable oil, included with soybean oil, corn oil, peanut oil, sesame oil and camellia oil was studied by using near infrared spectroscopy. The methods of mahalanobis distance and self-organizing competitive neural networks were used to classify those collected sampleswhich can build discriminative models. Simultaneously, the effects of the wavelength range and methods of spectra pretreatments on the models were explored. Both methods provided. To predict the validation set of 25 samples, the use of the established model of mahalanobis distance and self-organizing competitive neural networks showed very good discrimination between the oil classes with low classification error, the accuracy rate of which were reached to 100%.3. Quantitative determination of acid value and peroxide value were researched by the buildng of partial least squares (PLS) regression model and PLS-Back Propagation artificital neural network combined with near infrared spectroscopy,and the models were used to predict the validation set of 15 samples, respectively. The coefficient of determination of PLS model for acid value and peroxide value were 0.9837,0.9752, and the root mean square error of prediction (RMSEP) of 0.0752,0.00972, respectively. By using the models of BP network to predict the validation set of 15 samples the coefficient of determination R2 for acid value and peroxide value can reached to 0.9695,0.9744with the RMSEP of 0.595,0.00991 respectively. These two models can satisfy with the detected need of acid value and peroxide value, simultaneously.4. Qualitative and Quantitative determination of the adulterated binary system which was comprised of pure camellia oil samples mixed with various concentrations of soybean oil, corn oil and sunflower seed oil were researched by using near infrared spectroscopy. The discriminative model of between pure camellia oil and camellia oil adulterated, and model of amang three camellia oil samples adulterated were estabulished by mahalanobis distance cluster analysis. The accuracy rate useing the established models reached more than 99.1%. The quantitative determination models to detect concentrations of soybean oil, corn oil, and sunflower seed oil which adulterated into camellia oil binary system by partial least squares were established.The correlation coefficients of PLS were 0.99957,0.99962,0.99975, and RMSEC were 0.300,0.309,0.255. The RMSEP were 0.467,0.272, and 0.410, respectively. The true value and the predictive value of the validation set were compared by using paired t-test, the result showed that it was no significant difference and the research can obtained satisfactory results.5. The peanut oil adulterated binary system mixed with various concentrations of soybean oil, rapeseed oil and palm oil were prepared, and then the qualitative and quantitative determination of the system were studied by back propagation artificital neural network combined with near infrared spectroscopy. The BPnet could be divided into four classes processing 744 epochs. The accurate rate of the established BP neural network model to predict calibration set and validation set were reached to 100% and 96.0%, respectively. At the same time, by using BP-neural network, the quantitative determination models which were used to detect the concentrations of soybean oil, rapeseed oil, and palm oil in the peanut oil adulterated binary system were established, the predictive value of the validation set were obtain, also. The coefficients of determination were 0.9851,0.9901,0.9850, and the RMSEP were 1.05,1.10 and 1.72, respectively. Moreover, the results of the model were compared with the PLS and PCR, which show that PLS prediction model is better than the BP network, and both models can meet the needs of detecting the adulteration of peanut oil.6. The sesame oil samples adulterated binary system mixed with various concentrations of soybean oil, rapeseed oil and peanut oil were prepared. Qualitative determination of adulterate peanut oil was studied by Self-Organizing feature Map (SOM) neural network combined with near infrared spectroscopy. The spectrum datas were compressed by PCA to obtain scores of principal components. The optimal result was obtain when training epochs were reach to approximately 500 epochs with the accurate rate of 100%, and the accuracy rate of validation set 50 samples was reached to 95.6%. Subsequently, the quantitative determination models used to detect concentrations of soybean oil, rapeseed oil, and palm oil in peanut oil adulterated binary system were established by PLS. The RMSEP for validation set using PLS models were 1.05,1.31,0.686, and the R2 were 0.9951,0.9825,0.9944; separately. In this chapter, we also studied sesame oil samples adulterated ternary system mixed with varying concentrations of soybean oil and peanut oil, and the quantitative determination models were established to detect concentrations of soybean oil and peanut oil by partial least squares. The correlation coefficients of models were 0.95714 and 0.97025; The RMSEC were 1.54 and 1.42, respectively. The correlation is poor, which revealed that the method could not get an accurate.quantitation analyzed result.更多还原