【作者】 洪雪珍；

【导师】 王俊；

【作者基本信息】 浙江大学 ， 生物系统工程， 2014， 博士

【摘要】 近年来,消费者越来越青睐高品质、无添加、新鲜营养和少加工破坏的食物。鲜榨果汁的出现,正好符合消费者这个需求。本研究以樱桃番茄汁为对象,采用电子鼻和电子舌技术对樱桃番茄从榨汁前整果新鲜度、榨汁前处理、制汁、不同品质果汁混合、灭菌后放置的整个流程进行检测和评估,鉴定不同原料新鲜度、不同前处理方式、榨汁灭菌后放置不同时间和不同混合比例的樱桃番茄汁品质。本文从探索樱桃番茄汁的电子鼻和电子舌数据结构出发,比较基于单独电子鼻和电子舌的樱桃番茄汁品质鉴别方法的性能,然后利用数据融合方法优化樱桃番茄汁品质鉴别方法,最后提出半监督分类算法改善樱桃番茄汁品质分类模型。主要结论如下：(1)采用谱聚类、SL聚类法(Single Linkage clustering, SL)、CL聚类法(Complete Linkage clustering, CL)、Ward’s聚类法、FCM聚类法(Fuzzy c-means, FCM)、κ-means聚类法和ISODATA聚类法(Iterative Selforganizing Data Analysis Techniques Algorithm, ISODATA),分别对3个电子鼻和电子舌数据集底层结构进行探索。2维主成分分析(Principal Components Analysis, PCA)图和3个聚类有效性评价指标(Precision, Mutual Information和Rand Index)都表明,本研究提出的谱聚类算法能比传统的聚类法更好地反映出真实数据结构。(2)采用直接电子鼻检测、水汽处理后再电子鼻检测和电子舌检测,对4组不同混合比例的樱桃番茄汁品质进行鉴定。PCA、CDA (Canonical Discriminant Analysis, CDA)、LVQ (Learning Vector Quantization, LVQ), SVM (Support Vector Machines, SVM)和PCR (Principal Component Regression, PCR)结果表明,水汽处理后再进行电子鼻检测并不能提高对4组樱桃番茄汁的区分,而单独电子鼻或电子舌基本可以实现对混合果汁的分类和品质指标预测。(3)分别采用电子鼻和电子舌对不同原果新鲜度(4℃下储藏16天,25℃下储藏8天)的樱桃番茄汁进行检测。结果表明,单独电子鼻或电子舌可以辨识不同原果新鲜度的果汁,且基于电子舌数据建立的不同原果新鲜度果汁的分类和理化指标预测模型,都比基于电子鼻的好。(4)采用6种数据融合方式(直接融合法、PCA降维融合法、不同阈值的factor F选择法和逐步选择法)建立原果新鲜度追溯模型。结果表明,基于单个仪器建立的回归模型都缺乏泛化性,而数据融合后则可以对验证集样本的理化指标进行很好地预测。(5)比较cluster-then-label半监督分类法和有监督的PNN (Probabilistic Neural Networks, PNN)与LVQ建立的分类器性能。结果发现,对于不同原果新鲜度和不同混合比例的樱桃蕃茄汁数据集,基于谱聚类和FCM的Cluster-then-Label法的测试集正确率都高于两种有监督分类法,且半监督分类器对已知训练集和未知测试集的分类正确率差不多,其模型更稳定更具有泛化性。通过以上结论可知,利用电子鼻和电子舌技术对果汁品质进行鉴别是可行的。选用谱聚类算法可以更好了解果汁的电子鼻和电子舌数据集结构,且基于谱聚类的Cluster-then-Label半监督分类法能在只有少量有标签数据的情况下建立可靠分类模型。同时,当选择合适的特征提取方法后,数据融合比单独使用电子鼻和电子舌都要好。更多还原

【Abstract】 Nowadays, consumers demand high-quality, additive-free, minimally-processed, nutritious, and fresh-like products. Freshly-squeezed fruit juice labeled as100%fruit is typically one of these products. This research employed an electronic nose (e-nose) and an electronic tongue (e-tongue) to detect the quality of cherry tomato juices. The tasks of tracing material freshness, discriminating pretreatments, evaluating sterilization persistence and indenfying adulteration levels were considered. The whole structure of this paper is as follows:(a) explore data structure of the e-nose and e-tongue datasets of cherry tomato juices;(b) establish quality evaluation models for cherry tomato juices based on the e-nose and e-tongue detections, respectively;(c) optimize those quality evaluation models by combining the e-nose and e-tongue signals; and (d) present a semi-supervised approach to improve the quality discrimination classifiers. The main conclusions of this research are as follows:(1) Spectral clustering, single Linkage clustering (SL), complete Linkage clustering (CL), Ward’s clustering, Fuzzy c-means (FCM), κ-means clustering and iterative selforganizing data analysis techniques algorithm (ISODATA) were employed to explore the underlying data structure of three e-nose and three e-tongue datasets. Both2D principal components analysis (PCA) plots and three cluster validation criteria (CVC)-precision, mutual information (MI) and rand index (RI)-demonstrated that the spectral clustering outperformed traditional clustering methods at reflecting the real underlying structure of datasets.(2) The e-nose and e-tongue were applied to authenticate cherry tomato juices with different adulteration levels. In addition to directly e-nose measurement, a pretreatment of employing anhydrous sodium carbonate as desiccant was also conducted to observe if reducing of water vapor would improve the authentication ability of e-nose. The results of PCA, canonical discriminant analysis (CDA), learning vector quantization (LVQ), support vector machines (SVM) and principal component regression (PCR) all demonstrate that employing anhydrous sodium carbonate as desiccant did not improve the performance of e-nose when detecting liquid samples; on the contrary, directly e-nose measurement was better than e-nose with desiccant at authenticating cherry tomato juices. Meanwhile, either the e-tongue or the direct e-nose could authenticate cherry tomato juices as well as predict quality indices (pH and solube solid contents (SSC)).(3) The e-nose and e-tongue were applied to measure the freshness of cherry tomato under different storage conditions (16days at4℃and8days at25℃). Both qualitative and quantitative results demonstrate it is possible to trace the freshness of original cherry tomatoes through detecting the squeezed juices. Meanwhile, freshness discrimination models as well as quality indices prediction models built based on the e-tongue were better than those based on the e-nose.(4) A combination of the e-nose and e-tongue was tried to trace the freshness of cherry tomatoes by detecting the squeezed juices. Six data fusion approaches-simple concatenation, stepwise selection, PCA, factor F with Log F values higher than3,2.5and2, respectively-were employed. A second batch of experiments was conducted to produce an independent dataset for validation of the material freshness tracing models. The results demonstrated that quality regression models built based on single usage of e-nose or e-tongue were lack of generalization. On the contrary, models built based on fusion datasets could predict quality indices (pH, SSC, vitamin C and firmness) for the second batch of experimental data.(5) A semi-supervised classification approach-cluster-then-label-was compared with supervised approaches-probabilistic neural networks (PNN) and LVQ. In both cases of tracing material freshness and authenticating adulteration levels, cherry tomato juice quality evaluation models built using spectral clustering and FCM based cluster-then-label approaches outperformed the supervised approaches. In the cases that only a few labeled data is available for training, supervised classifiers tend to overfit the particular training dataset and thus lacking generalization ability, i.e., it tends to fail to classify unknown data. On the contrary, the semi-supervised classifier is more stable and of generalization. Classification accuracy for the training and testing datasets were almost the same.In consideration of the above conclusions, it is obvious that establishing quality evaluation models for cherry tomato juices based on the e-nose and e-tongue techniques is possible. The proposed spectral clustering offers a better approach to explore underlying structures of e-nose and e-tongue datasets; and based on this clustering approach, the semi-supervised cluster-then-lable approach could build a reliable classifier with only a few labeled training data. Meanwhile, with appropriate feacture extraction approaches, the fusion use of both instruements could be better than individual usage of the e-nose and e-tongue.更多还原