【作者】 王茜；

【导师】 孟庆浩； 王伟；

【作者基本信息】 天津大学 ， 控制工程， 2021， 硕士

【摘要】 酒是人类的主要饮料之一,酒文化在中国传统文化中有着独特的地位。在众多酒类中,白酒是我国人民的重要日常消费品。我国白酒品种丰富,不同的原材料、产地及工艺造就了不同香型的白酒。随着白酒需求的增大,不时出现质量安全问题,因此亟需快速、准确的白酒鉴别方法。现有的白酒检测通常通过人工嗅辨或分析仪器完成,前者主观差异性较大,后者费用高且时间长。近些年,电子鼻技术因速度快、成本低等优势开始在白酒检测领域发挥作用。电子鼻是一种模拟生物嗅觉系统的仿生气味检测仪器,可以实现对混合气体的快速识别。本文选用实验室自制的手持式电子鼻系统,研究不同品牌和香型白酒的识别方法,将重点聚焦在电子鼻传感器阵列输出信号的特征提取与降维两个方面。本文完成的主要研究工作如下:(1)提出了多域特征加权融合方法。首先,采用统计学分析、小波包分析和一对多共同空间模式(one versus rest common spatial pattern,OVR-CSP)方法分别提取白酒样本的时域、时频域和空域特征,其次,利用特征加权的方法融合多域特征,得到更全面的表征白酒样本的数据,用于后续分析和处理。(2)提出了一种新的有监督的特征降维方法L1-KECA-LDA。该方法将基于L1范数的核熵成分分析(L1-norm-based kernel entropy component analysis,L1-KECA)和线性判别分析(linear discriminant analysis,LDA)方法相结合,充分利用白酒样本的标签信息,得到更利于分类的约简特征。利用粒子群优化(particle swarm optimization,PSO)算法分别寻找KECA、L1-KECA和L1-KECA-LDA在支持向量机(support vector machine,SVM)、k近邻(K-nearest neighbor,KNN)和极限学习机(extreme learning machine,ELM)三种分类器下的最优高斯核参数。(3)将SVM、KNN和ELM三种分类器分别与所提出的特征提取和特征降维算法相结合用于白酒分类。分别比较了单一域特征与多域融合特征两种提取方法,KECA、L1-KECA和L1-KECA-LDA三种特征降维方法,以及SVM、KNN和ELM三种分类器对识别性能的影响。实验结果表明,当使用多域特征加权融合提取算法,配合L1-KECA-LDA降维以及ELM分类时,可以获得最好的白酒识别效果。更多还原

【Abstract】 Liquor is one of the main drinks of mankind,and liquor culture has a unique position in Chinese traditional culture.Among many alcoholic beverages,liquor is an important daily consumer goods for Chinese people.Chinese liquor is rich in variety,and different raw materials,producing areas and technologies have created different flavors of liquors.With the increasing demand for liquor,quality and safety problems occur from time to time.Therefore,rapid and accurate liquor identification methods are urgently needed.Currently,the detection of liquor is usually done by manual sniffing or analytical instruments,the former has a large subjective difference,and the latter is costly and time-consuming.In recent years,electronic nose(e-nose)technology has been playing a role in the field of liquor detection due to its advantages of high speed and low cost.E-nose is a kind of bionic odor detection instrument which simulates the biological olfactory system and can realize the rapid recognition of mixed gases.This thesis studied the identification methods of different brands and flavors of liquor with the self-made hand-held e-nose system and focused on the feature extraction and dimensionality reduction of the output signal of the e-nose sensor array.The main research work completed is as follows:A multi-domain feature weighted fusion method was proposed.First,statistical analysis,wavelet packet analysis,and one versus rest common spatial pattern(OVRCSP)methods were used respectively to extract the time domain,time-frequency domain,and spatial domain features of the liquor samples.Secondly,the multi-domain features are fused using a feature weighted method to obtain more comprehensive representation of the data of liquor samples for subsequent analysis and processing.A new supervised feature dimensionality reduction method named L1-KECALDA was proposed.This method combined L1-norm-based kernel entropy component analysis(L1-KECA)and linear discriminant analysis(LDA)methods to make full use of the label information of liquor samples for obtaining better reduction features for classification.Partial swarm optimization(PSO)algorithm was used to find the optimal Gaussian kernel parameters of KECA,L1-KECA and L1-KECA-LDA under the three classifiers of support vector machine(SVM),k-nearest neighbor(KNN)and extreme learning machine(ELM).Three classifiers,SVM,KNN and ELM,were combined with the proposed feature extraction and dimensionality reduction algorithms for liquor classification.Two extraction methods of single-domain feature and multi-domain fusion feature,three feature dimensionality reduction methods of KECA,L1-KECA and L1-KECA-LDA,and three classifiers of SVM,KNN and ELM were compared on the recognition performance.The experimental results show that when using the multi-domain feature weighted fusion extraction algorithm,combined with L1-KECA-LDA dimensionality reduction and ELM classification,the best liquor recognition effect can be obtained.更多还原