【作者】 陈燕清；

【导师】 倪永年；

【作者基本信息】 南昌大学 ， 食品科学， 2010， 博士

【摘要】 食品真实性鉴别和食品中的危害物质分析是食品质量与安全控制中两个重要的研究内容,关系到消费者切身利益和身体健康。研究快速有效的食品质量鉴别、分类和有毒有害物质分析技术具有重要的现实意义。化学计量学运用数学、统计学、计算机科学及其它相关科学的理论与方法,优化化学测量过程,分辨复杂波谱,最大限度从化学测量数据中获得有用的化学分类信息,是一门化学测量的理论与方法学。将化学计量学结合现代分析技术应用在食品质量与安全控制中,为食品安全与检测提供了新方法。本论文共分为六章。主要研究了将化学计量学模式识别方法结合可见-红外、原子吸收、同步荧光等分析技术,应用于不同种类和产地的酱油、食醋和料酒的分类鉴别；将多元校正技术结合紫外-可见分光光度法应用于食品防腐剂多组分重叠光谱解析和同时测定。第一章介绍了食品质量研究中常用的一些化学模式识别方法(聚类分析、主成分分析、判别分析)的基本原理,并介绍了模式识别结合红外、原子吸收、原子发射、气相色谱、液相色谱、质谱、传感器等检测技术在食品质量控制中的应用；介绍了多元校正技术(偏最小二乘、主成分回归、平行因子等)和人工神经网络在食品分析中的原理及应用,指出了化学计量学方法在食品质量控制和食品分析中的应用前景。第二章研究了以酱油的9个理化指标为变量,采用模式识别方法对不同种类和品牌酱油进行分类和质量鉴别。收集了三个不同品牌的53个酱油样品,其中26个生抽,27个老抽样品。通过化学方法测定了酱油的9个理化参数值(密度、酸度、总固形物、灰分、电导、氨基酸氮、食盐、粘度及总酸)作为酱油样品的特征变量。采用夹角余弦法计算了不同样品之间的相似度,评价产品质量的稳定信息和不同品牌样品的区分度。结果表明相似度法对判断酱油质量的的稳定性有一定作用,而对酱油品牌的区分有一定的局限性。聚类分析和主成分分析探讨不同品牌和种类酱油区分的可行性,结果显示不同品牌和种类的酱油能各自聚在一起,表明本研究所选择的变量的有效性。分别采用偏最小二乘、线性判别和K-最邻近法三种判别模型对预报集中酱油的品牌和种类进行判断,结果表明三种模型均能很好的判断酱油的品牌和种类。在建立线性判别和K-最邻近法判别模型前,采用Fisher权重法计算不同变量对酱油分类的贡献大小,采用交叉验证法计算变量个数对线性判别判断准确率关系曲线,得到前7个贡献大的变量对预报集中酱油的品牌和种类判断的准确率能达到100%,即选择密度、固形物、总酸、pH、氯化钠、电导、灰分的测量数据建立数据矩阵。第三章研究分别以食醋中8种微量元素含量值和5个理化指标值为变量,探讨模式识别方法对不同种类和产地食醋进行分类鉴别的可行性。实验购买4个不同品牌29个食醋样品,包括陈醋和白醋。采用原子吸收分光光度法测食醋中8种微量元素含量,化学方法测定食醋5个理化指标值,组成测量数据矩阵。采用向量相似法计算不同品牌和种类食醋的质量稳定性信息,研究两种不同类型变量(金属元素含量和理化指标值)对食醋的区分效果。主成分分析分别用理化指标、金属元素含量值及理化指标和金属元素含量的混合数据为变量,以其得出两种不同类型变量对食醋种类和品牌区分的贡献。从13个变量的载荷图可以得出,金属元素为变量对7个样品类别区分的贡献比5个理化指标的贡献大。在由PC1-PC2-PC3前三个主成分构成的三维得分图中,32个样本按种类和品牌被成功的区分为7组。采用夹角余弦计算样品的距离,32个样本能得到很好的聚类效果。采用建立的偏最小二乘和径向基人工神经网络判别模型分别对测试集进行种类预报,预报准确度分别达到100%和93%。第四章测定不同品牌料酒的可见-近红外光谱,建立模式识别方法对料酒品牌区分的新方法。实验购买了3个不同品牌共37个料酒样品,测定400-1400nm波长范围的可见近红外的吸收光谱,分别采用一阶导数法和小波变换技术对料酒的可见-近红外光谱数据去噪和压缩处理。探讨了小波分解尺度对光谱信号的影响,最后选择二阶的Daubechies (db2)小波函数、分解尺度5对原始光谱数据进行处理。对原始光谱数据、一阶导数法和小波变换技术处理后的数据进行主成分分析,比较了模糊聚类的效果。结果表明,小波变换能够有效去除光谱噪音和压缩光谱变量,得到较好的聚类效果。采用偏最小二乘和人工神经网络预报模型对料酒的品牌进行判断,预报正确率均为100%。第五章测定料酒的三维同步荧光,提取特征荧光变量,建立模式识别方法对料酒品牌区分的新方法。分别采用主成分分析降维和小波变换的方法提取三维荧光的特征变量。主成分分析取第一主成分作为料酒的荧光特征变量。通过比较小波分解尺度对光谱信号的影响的结果,最后选择二阶的Daubechies(db2)小波分解尺度4对原始的光谱数据进行处理,该方法能很好的压缩和保留原始的荧光信息。对主成分降维和小波变换的方法提取三维荧光的特征变量进行主成分分析和聚类分析,小波变换法能更好的对样品品牌进行区分。采用偏最小二乘和人工神经网络预报模型对料酒的品牌进行判断,预报正确率均为100%。第六章研究了多元校正技术和人工神经网络等化学计量学方法解析光谱严重重叠的苯甲酸、对羟基苯甲酸甲酯、对羟基苯甲酸丙酯和山梨酸四种防腐剂的紫外吸收光谱,建立了同时测定四种防腐剂的新方法。考察酸度对吸收光谱的影响,发现在酸性溶液中,四种防腐剂的测定灵敏度较高。选择在pH 2.09的B-R缓冲溶液中,对四种防腐剂进行同时测定。在优化的酸度条件下苯甲酸、对羟基苯甲酸甲酯、对羟基苯甲酸丙酯单组分的线性范围为0.5-20μg mL-1,山梨的线性范围为0.25～10μg mL-1。四种防腐剂的检测限分别为0.22μg mL-1,0.19μg mL-1,0.17μg mL-1,0.085μg mL-1。采用多种校正模型(经典最小二乘(CLS)、偏最小二乘(PLS)、主成分回归(PCR)、一阶导经典最小二乘(DCLS)、一阶导偏最小二乘(DPLS)、一阶导主成分回归(DPCR)及径向基人工神经网络(RBF-ANN)对四组分的合成样预报集浓度预报。结果表明,建立的校正模型均有较好的预报能力,相对预报误差(RPET)小于10%。其中,PCR、DPCR和RBF-ANN的预报误差相对较小,预报误差分别为4.53%,4.55%和4.67%。用建立的PCR和RBF-ANN模型结合光度法对实际样品中四种防腐剂直接同时测定,获得满意结果。更多还原

【Abstract】 Discrimination of food authenticity and determination of toxic and harmful matter are two crucial issues in food safety and quality control. It is related to the consumers’interests and health. It is realistic significance to develop rapid and effective method for discrimination of food authenticity and determination of toxic and harmful materials.Chemometrics is an object on chemical theory of measurement and methodology. It can optimize the process of measurement, resolve overlapped spectra and extract maximum useful discrimination information of data from chemical measurement using mathematics, statistics, computer science and other related science theories and methods. It offers a new method to solve some problems about food safety and food quality control by using chemometrics and modern analytical means.There are six chapters in this thesis. The focuses of the research work are discrimination of soy sauce samples, vinegar samples and seasoning wine samples of different brands and kinds by NIR, AAS and SFS with the aid of chemical pattern recognition techniques, and resolution of overlapped spectra of four preservatives and determination simultaneously by UV-visible spectrometry with the aid of multivariate calibration.Chapter one The principals of some chemical pattern recognition and multivariate calibration techniques (CA, PCA, DA, PLS, PCR, ANN and PARAFAC), and application of IR, AAS, AES, GC, HPLC and MS in food safety and quality control combined with chemical pattern recognition techniques and in food analysis combined with multivariate calibration were reviewed and summarized. The outlook of chemometrics in food quality and food control was also discussed.Chapter two A new method of discrimination of soy sauce samples of different kinds and brands was developed according to 9 physico-chemical variables using pattern recognition.53 soy sauce samples of three different brands were collected including 26 light soy sauces and 27 dark soy sauces. The values of 9 physico-chemical properties (density, pH, dry matter, ashes, electro-conductivity, amino-nitrogen, salt and total acidity) were determined and acted as the characteristic variables of soy sauce samples. To evaluate the stability of the quality and degree of differentiation of different brands, the similarities of different products were calculated by vector similarity analysis. The results showed that SA was useful to evaluate the stability of soy sauce quality, but limited to differentiate the brands and kinds of soy sauce samples. We used cluster analysis and principal component analysis to study whether it was feasible to discriminate the brands and kinds of soy sauce samples. The results of cluster analysis and principal component analysis showed the effectiveness of discrimination and the correctness of variables selected to predict the brands and kinds of soy sauce samples in verification set. In order to predict the unknown samples, several calibration models were set up, such as partial least squares, linear discrimination analysis and K-nearest neighbor. The results of prediction showed that these models were effective to discriminate soy sauce samples. The variables for LDA and KNN were chosen by means of Fisher F-ratio approach, and the prediction ability of all classifier was evaluated by cross-validation. The first seven variables (density, dry matter, total acidity, pH, salt, electro-conductivity and ashes) were chosen according to the curve of the numbers of variables and correct classification rates. Among the three supervised discrimination analysis, LDA and KNN gave 100% predications according to the categories and brands of samples.Chapter three The research discussed the feasibility of discrimination of vinegar samples of different kinds and brands according to 8 metallic contents and 5 physico-chemical parameters with the aid of pattern recognition techniques.29 vinegar samples, including mature vinegar and white vinegar, were collected. The metallic contents were determined by AAS, and physico-chemical parameters were determined by chemical methods. The data measured were acted as characteristic variables of vinegar samples. To evaluate the stability of vinegar quality similarities of different products were calculated by SA. In order to comparing the contribution of the two kinds of data (metallic contents and physico-chemical parameters) to the discrimination of vinegar samples of different kinds and brands, they were used as variables for principal component analysis, respectively. The loading plot of 13 variables showed that metallic contents contributed greater than physico-chemical parameters.32 vinegar samples were divided into seven groups according to the kinds and brands in the three-dimensional space of the first three PCs.32 vingar samples were correctly clustered according to the distance calculated using Angle Cosine function. The correct prediction rate were 100% and 93% for verification set by PLS model and RBF-ANN models, respectively.Chapter four A new method of fast discrimination of brands of seasoning wine by means of visible-near infrared spectroscopy was developed. The visible-near infrared absorption spectroscopic signals of 37 seasoning wine samples from three different brands were measured between 400 nm and 1400 nm. The spectroscopic data were pretreated by first derivative and WT to denoise and compress data, and the impact of the level of wavelet decomposition on the original spectra was also discussed. In this study we employed second order Daubechies (db2) wavelet function and the fifth decomposition level to denoise and compress the original data. The results of PCA were compared by using original data, data treated by first derivative and WT as characteristic variables. From the clear classification result by PCA and CA, we showed that WT can denoise and compress data effectively. PLS and RBF-ANN calibration models were used to predict the brands of seasoning wine in verification set with 100% accuracy of prediction.Chapter five A new method for fast discrimination of brands of seasoning wine by using characteristic variables extracted from three dimensional SFS was developed. The original three dimensional fluorescence data was compressed and extracted by PCA and WT. The first PC was used as the characteristic fluorescence variables of seasoning wine samples by PCA. By comparing the effects on the signals of the different levels of decomposition, db2 wavelet function and the fourth decomposition level were chosen to extract and compress the data to obtain the original characteristic signal information. From the results of PCA and CA using the characteristic variables extracted by PCA and WT, brands of seasoning wine sample were correctly classified more by WT. The two supervised discrimination analysis calibration models, PLS and RBF-ANN gave 100% predications for unknown samples in the verification set according to the brands of seasoning wine samples. Chapter six Benzoic acid (BA), methylparaben (MP), propylparaben (PP) and sorbic acid (SA) are food preservatives, and they have well defined UV spectra. However, their spectra overlap seriously, and it is difficult to determine them individually from their mixtures without preseparation. The multivariate calibration and RBF-ANN of chemometrics were applied to resolve the overlapping spectra and to determine these compounds simultaneously. The influence of acidity on absorption spectra was investigated. It was discovered that in the acidic buffer solution the sensitivity of detection was higher than in basic buffer solution. Therefore, determination of four preservatives was conducted in pH 2.09 B-R buffer solution. Under the optimum acidic condition, the four compounds, when taken individually, behaved linearly in the 0.25-20 mg L-1 for BA, MP, PP and 0.25-10 mg L-1 for SA concentration range, and the limits of detection (LOD) were 0.22,0.19,0.17 and 0.085 mg L-1 for BA, MP, PP and SA, respectively. Multivariate calibration (CLS, PCR, PLS, DCLS, DPCR, DPLS) and RBF-ANN models were applied to predict the concentration of the four preservatives in the verification set. The results of prediction showed that the calibration models were effective to correctly predict the individual concentration in the mixture, and relative prediction errors (RPET) were under 10%. Among those models, PCR, DPCR and RBF-ANN gave more satisfactory results, and the RPETS were 4.53%,4.55% and 4.67%, respectively. It was obtained satisfactory results to determine the four preservatives simultaneously by UV-visible spectrometry with the aid of chemometrics.更多还原