葡萄籽油掺假种类繁多, 手段隐蔽, 成为食品安全检测的重要难点之一, 为规范食用油市场, 提供一种方便、 可靠的葡萄籽油品质鉴别方法尤为重要。 针对色谱和质谱等传统品质分析方法的耗时、 试剂消耗大、 专业性强等不足, 以及实现无损分析的近红外光谱仪价格昂贵、 操作环境要求高等缺点, 研究设计一套低成本、 高准确度的可见/近红外光谱仪检测系统来实现葡萄籽油品质掺假鉴别。 首先, 依托USB6500-Pro探测器搭建可见/近红外光谱仪硬件平台, 并基于Qt设计一套简洁的人机交互界面, 用以实现光谱数据的采集、 处理以及葡萄籽油掺假鉴别结果的显示; 其次, 针对硬件和检测环境带来的光谱噪声, 系统采用小波变换滤除噪声, 减小光谱失真; 最后, 考虑到现有的基于机器学习的品质鉴别模型往往依赖已知的油类训练样本集来实现对不同掺假类别油类的预测, 而利益驱使下层出不穷的掺假手段使得新的、 未出现在原训练集中的掺假类别样本不断涌现, 现有的品质鉴别方法将很难给出准确的判别结果。 因此, 研发的检测系统中设计一种能实现已知和新的掺假油品光谱的鉴别方法, 该方法分为分类和校正两步: 先用建模数据库中的训练集建立极限学习机(ELM)分类器模型, 实现初步掺假类别的分类; 然后再利用自动聚类算法对分类结果进一步校正。 若与校正数据集产生一个聚类中心, 则证明分类结果正确且属于建模数据库中的已知掺假类别; 若产生两个聚类中心, 则分类结果不正确, 样本为新掺假类别, 未出现在建模数据库中, 最终得到准确的掺假类别结果。 为检验系统性能, 用搭建的可见/近红外光谱仪硬件平台采集了纯葡萄籽油和掺入不同比例的大豆油、 玉米油、 葵花籽油和调和油的葡萄籽油的5类光谱数据, 每一类30组共计150组数据, 将得到的可见/近红外光谱数据先进行小波阈值法去噪和多元散射校正(MSC)预处理后输入到设计的检测系统中。 假定前4类作为建模数据库中的已知掺假类别以及第5类作为新掺假类别, 先利用K-S算法将已知掺假类别的每类样本划分成训练集20组和测试集10组, 用训练集共80组样本建立ELM分类模型, 将40组测试集输入到ELM实现初步判别, 判别结果再进一步聚类分析校正, 只有一个聚类中心, 说明了模型判别准确, 且对已知类别能够100%识别; 当30组新掺假类别样本输入到ELM模型时, 均判别成了纯葡萄籽油, 进一步聚类分析校正, 产生了两个聚类中心点, 说明ELM模型误判, 定性判定第5类为新掺假类别。 实验结果表明, 研发的葡萄籽油掺伪检测系统操作简单、 快速, 不仅对已知掺假类别能够100%识别, 而且对新掺假类别能够实现定性判别。

Various kinds of adulterated grape seed oil and concealed adulterated means cause a severe problem in food safety detection. In order to regulate the edible oil market, it is especially important to provide a convenient and reliable method for identifying the quality of grape seed oil. However, traditional methods for chromatography and mass spectrometry are time consuming, reagent intensive, highly specialized, etc.; and the near infrared spectrometer that realizes non-destructive analysis is expensive and has high operating environment requirements. Thus, a visible/near infrared spectrometer with low cost and high accuracy was designed to discriminate grape seed oil adulteration. Firstly, a visible/near infrared spectrometer hardware platform based on USB6500-Pro detector was built, and a simple human-computer interaction interface based on Qt was designed to realize the collection and processing of spectral data and the display of grape seed oil adulteration discrimination results. Secondly, for the spectral noise brought by hardware and detection environment, wavelet transform was used to filter out noise and reduce spectral distortion. Finally, considering that the existing quality discrimination models based on machine learning often rely on the known oil training sample set to predict the different adulterated categories; and driven by interest adulteration means will emerge in endlessly which will result in the emerging of new adulteration categories not in the original training set, the existing quality identification methods are difficult to give accurate results. Therefore, a discrimination method for known and new adulterated oil spectra was designed in the detection system. This method was realized by two steps: (1) classification: the extreme learning machine (ELM) classifier model was established by using the training set in the modeling database to realize the preliminary judgment of the preliminary adulteration category; (2) correction: the automatic clustering algorithm was then used to further correct the prediction result. If a clustering center is generated with the correction data set, it is proved that the prediction result is correct and belongs to the known adulteration category in the modeling database; if two cluster centers are generated, the prediction result is incorrect and the sample is a new adulteration category which does not appear in the modeling database. The result of the accurate adulterated category was eventually obtained. In order to test the performance of the system, five classes of oil, including pure grape seed oil, and grape seed oil blended with different proportions of soybean oil, corn oil, sunflower oil and blend oil were analyzed by the visible/near infrared hardware platform and their spectroscopy data were collected. It contains 30 sets of data for each class of oil, totals 150 sets. Before inputting the visible/near infrared spectroscopy data into the detection system, they were firstly de-noised by wavelet threshold method and pre-processed by multiple scattering correction. Assuming that the first four classes were known adulteration class in the modeling database and the fifth class was new adulteration class, samples from each of the four known adulteration classes were divided into 20 training sets and 10 test sets by using K-S algorithm. Then, ELM classification model was established by using 80 training sets, and 40 test sets were input into ELM for preliminary discrimination. The discrimination results were further analyzed and corrected by clustering. There was one clustering center, which meant that the ELM model discriminated accurately and could recognize 100% of the known classes. However, when 30 samples from the new adulterated class were put into the ELM model, all of them were discriminated as pure grape seed oil. The discrimination results were further clustered and corrected. There were two clustering centers, which showed that the model was misjudged and the fifth class was qualitatively determined as a new adulterated class. The experimental results showed that the designed visible/near infrared spectroscopy detection system was simple and fast, and can identify not only the known adulteration categories but also the new adulteration categories.