基于关联向量机的气体传感器阵列信号处理
发布时间：2018-10-23 15:22
【摘要】：气体传感器阵列由多个传感器构成,利用气体传感器特有的“交叉敏”特点,每个传感器对要测的气味信息都有着不同的敏感度、选择性和反复性。在传感器阵列多维空间中形成响应模式,阵列所确定的多维空间也包含了更多的信息。但是由于现在的气味成分信息、越来越复杂,仅靠传感器阵列已经不能全面和准确的分析气味信息,所以本文引入关联向量机(Relevance Vector Machine, RVM)对气体传感器阵列信号进行处理。RVM是一种新的机器学习方法,用于气体传感器阵列信号处理,具有良好的泛化性能、概率式预测等特点。本文选取中草药白薇和五种常见的稻米为例。采用传感器阵列、模式识别技术相结合的方法对中草药白薇货架期进行判定和检测稻米的质量。在进行实验样本分类之前,本文采用主成分分析法对数据进行预处理,不仅降低了计算的复杂程度,同时也提高了分类效率。为了突出本文RVM分类的有效性和可行性,将其和支持向量机(Support Vector Machine, SVM)、神经网络等方法进行对比。本论文的主要研究工作如下：(1)传感器阵列信号的采集。本文选择5种稻米和中草药白薇作为研究对象。使用电子鼻进行实验,采集实验样本信息。用主成分分析法对实验样本进行特征提取,并对初始特征向量进行主成分分析,保留样本的主要成分信息,达到降维的目的,减少计算量,提高分类效率。(2)采用RVM对实验样本进行分类和货架期判定。采用实验法确定分类模型的核函数及核参数,并比较不同核函数和核参数对分类识别精度的影响,从而确定最优的分类模型。在二分类实验中,实验结果表明采用高斯(Gauss)和三次多项式(Ploy3)核函数,分类精度较高,相关向量数较少,分类所需时间相对较短。选择Poly3核函数时运行时间最短,便于在线实时检测。再将RVM二分类推广到多分类,比较不同核函数和核参数时,分类模型的准确率。(3)比较不同算法的分类精度。对RVM、SVM和神经网络方法进行分类比较。实验结果验证了本文算法可以有效的克服SVM存在的测试时间长、支持向量个数过多等问题。从测试时间来看,在BP、RBF和SVM都选择高斯核函数时,RVM的测试时间比BP、RBF、SVM都要短很多。RVM不需要改动其他参数,核函数也不用符合Mercer条件。通过实验结果表明,将RVM的分类技术引入到稻米分类和货架期判定,可以保证较高的分类精度,获得的模型更加稀疏、测试时间也缩短。与其他方法的对比结果,验证了本文RVM模型用于分类识别的有效性和可行性,同时也适用于其他样本的分类检测。
[Abstract]:The gas sensor array is composed of multiple sensors. Each sensor has different sensitivity, selectivity and repetition to the odour information to be measured by using the characteristic of "cross sensitivity" of gas sensor. The response mode is formed in the sensor array multidimensional space, and the multidimensional space determined by the array also contains more information. However, due to the increasing complexity of odour information, it is not possible to analyze odor information comprehensively and accurately by using sensor arrays alone. In this paper, we introduce the correlation vector machine (Relevance Vector Machine, RVM) to process the gas sensor array signal. RVM is a new machine learning method, which is used in the gas sensor array signal processing, and has good generalization performance and probabilistic prediction. In this paper, Bai Wei and five common rice varieties are selected as examples. Using sensor array and pattern recognition technology, the shelf life of Chinese herbal medicine Bai Wei was determined and the quality of rice was tested. Before the experiment sample classification, the principal component analysis (PCA) is used to preprocess the data, which not only reduces the complexity of calculation, but also improves the classification efficiency. In order to highlight the validity and feasibility of RVM classification in this paper, it is compared with support vector machine (SVM) (Support Vector Machine, SVM), neural network. The main work of this thesis is as follows: (1) acquisition of sensor array signal. In this paper, five kinds of rice and Chinese herbal medicine Bai Wei were selected as the research object. The electronic nose is used to carry out the experiment, and the information of the experiment sample is collected. The principal component analysis (PCA) is used to extract the features of the experimental samples, and the initial eigenvector is analyzed to preserve the information of the main components of the samples, so as to achieve the purpose of reducing the dimension and reducing the amount of calculation. Improve the classification efficiency. (2) RVM was used to classify the experimental samples and determine the shelf life. The kernel function and kernel parameters of the classification model are determined by the experimental method, and the effects of different kernel functions and kernel parameters on the classification recognition accuracy are compared to determine the optimal classification model. In the two-classification experiment, the experimental results show that Gao Si (Gauss) and cubic polynomial (Ploy3) kernel functions have higher classification accuracy, fewer correlation vectors and shorter classification time. The Poly3 kernel function has the shortest running time and is easy to detect in real time. Then the RVM two-classification is extended to multi-classification to compare the accuracy of the classification models with different kernel functions and kernel parameters. (3) the classification accuracy of different algorithms is compared. The classification of RVM,SVM and neural network is compared. Experimental results show that the proposed algorithm can effectively overcome the problems of long test time and excessive number of support vectors in SVM. From the point of view of test time, when BP,RBF and SVM choose Gao Si kernel function, the test time of RVM is much shorter than that of BP,RBF,SVM. RVM does not need to change other parameters and the kernel function does not have to meet the Mercer condition. The experimental results show that the introduction of RVM classification technology to rice classification and shelf life determination can ensure higher classification accuracy, the obtained model is more sparse, and the test time is shortened. Compared with other methods, the effectiveness and feasibility of the proposed RVM model for classification and recognition are verified. At the same time, it is also applicable to the classification and detection of other samples.
【学位授予单位】：浙江师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP212;TN911.7
本文编号：2289644
[Abstract]:The gas sensor array is composed of multiple sensors. Each sensor has different sensitivity, selectivity and repetition to the odour information to be measured by using the characteristic of "cross sensitivity" of gas sensor. The response mode is formed in the sensor array multidimensional space, and the multidimensional space determined by the array also contains more information. However, due to the increasing complexity of odour information, it is not possible to analyze odor information comprehensively and accurately by using sensor arrays alone. In this paper, we introduce the correlation vector machine (Relevance Vector Machine, RVM) to process the gas sensor array signal. RVM is a new machine learning method, which is used in the gas sensor array signal processing, and has good generalization performance and probabilistic prediction. In this paper, Bai Wei and five common rice varieties are selected as examples. Using sensor array and pattern recognition technology, the shelf life of Chinese herbal medicine Bai Wei was determined and the quality of rice was tested. Before the experiment sample classification, the principal component analysis (PCA) is used to preprocess the data, which not only reduces the complexity of calculation, but also improves the classification efficiency. In order to highlight the validity and feasibility of RVM classification in this paper, it is compared with support vector machine (SVM) (Support Vector Machine, SVM), neural network. The main work of this thesis is as follows: (1) acquisition of sensor array signal. In this paper, five kinds of rice and Chinese herbal medicine Bai Wei were selected as the research object. The electronic nose is used to carry out the experiment, and the information of the experiment sample is collected. The principal component analysis (PCA) is used to extract the features of the experimental samples, and the initial eigenvector is analyzed to preserve the information of the main components of the samples, so as to achieve the purpose of reducing the dimension and reducing the amount of calculation. Improve the classification efficiency. (2) RVM was used to classify the experimental samples and determine the shelf life. The kernel function and kernel parameters of the classification model are determined by the experimental method, and the effects of different kernel functions and kernel parameters on the classification recognition accuracy are compared to determine the optimal classification model. In the two-classification experiment, the experimental results show that Gao Si (Gauss) and cubic polynomial (Ploy3) kernel functions have higher classification accuracy, fewer correlation vectors and shorter classification time. The Poly3 kernel function has the shortest running time and is easy to detect in real time. Then the RVM two-classification is extended to multi-classification to compare the accuracy of the classification models with different kernel functions and kernel parameters. (3) the classification accuracy of different algorithms is compared. The classification of RVM,SVM and neural network is compared. Experimental results show that the proposed algorithm can effectively overcome the problems of long test time and excessive number of support vectors in SVM. From the point of view of test time, when BP,RBF and SVM choose Gao Si kernel function, the test time of RVM is much shorter than that of BP,RBF,SVM. RVM does not need to change other parameters and the kernel function does not have to meet the Mercer condition. The experimental results show that the introduction of RVM classification technology to rice classification and shelf life determination can ensure higher classification accuracy, the obtained model is more sparse, and the test time is shortened. Compared with other methods, the effectiveness and feasibility of the proposed RVM model for classification and recognition are verified. At the same time, it is also applicable to the classification and detection of other samples.
【学位授予单位】：浙江师范大学
【学位级别】：硕士
【学位授予年份】：2014
【分类号】：TP212;TN911.7
【参考文献】
相关期刊论文 前1条
1 范伊红;李敏;张元;;相关向量机在车型识别中的应用研究[J];计算机工程与设计;2008年06期
，本文编号：2289644
本文链接：https://www.wllwen.com/kejilunwen/wltx/2289644.html
