傅里叶变换结合卷积神经网络分析西藏酥油真伪  

Fourier Transform Combined with Convolutional Neural Network to Analyze the Authenticity of Tibetan Ghee

作　　者：扎西穷达 次仁旺姆 央拉 尼珍 Zhaxiqiongda;Cirenwangmu;Yangla;Nizhen(Tibet Institutes for Food and Drug Control,Tibetan Autonomous Region Medical Ring Testing Center,Key Laboratory of Quality Control Center of Traditional Chinese Medicine(Tibetan Medicine)State Drug Administration of China,Lhasa 850000,China)

机构地区：[1]西藏自治区食品药品检验研究院,西藏自治区医疗器戒检测中心,国家药品监督管理局中药(藏药)质量控制中心重点实验室,西藏拉萨850000

出　　处：《现代食品》2023年第13期148-153,共6页Modern Food

基　　金：西藏自治区食品药品检验研究院一般项目“近红外光谱仪快速鉴别西藏酥油真伪的初步研究”(XZSYJY-YJKYXM-2020-3)。

摘　　要：近红外光谱是分析研究西藏酥油的重要方法。在时间、地域、温度等随机因素的影响下,实际谱图分析中会产生较多干扰,无法准确归纳研究其系统规律,只能从已有的随机信号中筛选建立研究模型。本文提出了一种通过傅里叶变换将近红外光谱时域信号转化为频域信号的方法,深入揭示近红外光谱的变化与西藏酥油性质的联系,使用一阶导数法对近红外光谱进行降噪处理,结合卷积神经网络判别西藏酥油的真伪。结果表明,经傅里叶变换后,近红外光谱维度大幅度降低至90%,在350 Hz、600 Hz附近发现了西藏酥油的特征光谱信息,处理后近红外光谱的特征性进一步显现,酥油鉴别模型的AUC值由79%提升至97%。在大数据集建模过程中,卷积神经网对真假酥油有良好的辨别能力,AUC值均为1。但由于样本数据量的限制,该方法在小数据集建模中的随机性加大,AUC值为0.5。Near infrared spectroscopy is an important method to analyze and study Tibetan ghee.Under the influence of random factors such as time,region and temperature,there will be more interference in the actual spectral analysis,and it is impossible to accurately conclude the research system law,so the research model can only be established by screening the existing random signals.In this paper,a method of converting near-infrared spectrum signal from time domain signal to frequency domain signal by Fourier transform is proposed to deeply reveal the relationship between the change of near-infrared spectrum and the property of Tibetan ghee.The first derivative method is used to reduce the noise of near-infrared spectrum,and the convolution neural network is combined to identify the authenticity of Tibetan ghee.The results showed that,after Fourier transform,the NIR spectral dimension was reduced to 90%,and the characteristic spectral information of Tibetan ghee was found near 350 Hz and 600 Hz.After processing,the characteristic spectral information of Tibetan ghee was further developed,and the AUC value of ghee identification model increased from 79%to 97%.In the large data set modeling process,the convolutional neural network has a good ability to distinguish true and false ghee,and the AUC value is 1.However,due to the limitation of sample data size,the randomness of this method is increased in the modeling of small data sets,and the AUC value is 0.5.

关 键 词：酥油 傅里叶变换 近红外光谱 光谱采集 模型建立 

分 类 号：TS252.7[轻工技术与工程—农产品加工及贮藏工程]