【作者】 傅均；

【导师】 李光；

【作者基本信息】 浙江大学 ， 生物医学工程， 2009， 博士

【摘要】 电子鼻是模拟生物嗅觉原理的仿生检测仪器,通常由交叉敏感的化学传感器阵列和合适的模式识别算法组成,自动完成对气味的定性或定量辨识。与专业嗅觉师和现代分析化学仪器相比,其简单、客观和廉价的优点,已经在许多应用领域引起了广泛重视。而生物嗅觉基础研究的成果与计算神经科学交融,又进一步丰富了人工嗅觉理论,促进了电子鼻的发展。本论文介绍和研究了一种嗅觉神经网络——KⅢ网络,并尝试将其引入到电子鼻的仿生信息处理中。本文首先通过数值分析方法对KⅢ网络的数学模型进行仿真,结果显示其动力学特性与生物嗅觉实验中观察到的EEG信号在某些方面具有相似性；又通过小世界网络理论和小波分解方法对KⅢ网络的结构进行分析,发现其拓扑性质与复杂生物系统的小世界网络具有相似性。从而在功能和结构上展示了KⅢ网络的仿生性能。然后研究了KⅢ网络在气敏传感器漂移补偿和对气味浓度影响的鲁棒性等人工嗅觉理论问题上的应用。通过作者自制的电子鼻系统对6种有机挥发性气体进行分类,结果表明KⅢ网络对存在传感器漂移及较大浓度偏差的样本仍然具有很好的识别率。本文还提出了KⅢ网络的时空模式处理概念,通过KⅢ网络的可扩充并行结构与其时间依赖运算过程,和电子鼻的传感器阵列结构与其对气体的时间响应特性相结合,将KⅢ网络“无缝”融入电子鼻系统,进一步模拟了生物嗅觉系统信息处理过程。本文还发展了KⅢ网络的模式识别理论,将它与当前较为流行的支持向量机相结合,构造出不同的级联分类器和集群分类器。并通过对4种茶叶和6个品牌的鲜奶和还原奶的识别,进一步探索KⅢ网络在电子鼻实际应用中的表现。更多还原

【Abstract】 Electronic nose is a kind of bionic instrument to mimic the mechanism of biological olfactory systems, which generally consists of an array of cross-sensitive chemsensors and an appropriate pattern recognition algorithm, capable of automatically detecting and discriminating odors. Due to the advantages of simply, objective and low-cost compared to human panels and conventional analytical instruments, electronic noses have been widely used in many areas. Moreover, resent basic research findings in biological olfaction combined with computational neuroscience promote its development both in methodology and application. This thesis studies an olfactory neural network entitled KⅢand introduces it to electronic nose community as a bionic information processing technique.Firstly, the numeral analysis of the KⅢnetwork is carried out. The simulation results show that the properties of its dynamics are much similar with that of EEG signals observed in biological experiments.The topological properties of the KⅢnetwork is also studied using the small-world network theory and wavelet decomposition method. The results show the KⅢnetwork with some small-world network characteristics, similar to which in complex biological systems.Then, the application of the KⅢnetwork to some theoretical problems in artificial olfaction is investigated, such as sensor drift counteraction and concentration influence elimination. Experimental data of six volatile organic compounds obtained through a home-made electronic nose set-up are processed by the KⅢnetwork. The results show that the KⅢhas good performances in classification of samples with sensor drift or concentration variation. In this thesis, a spatio-temporal pattern processing conception is proposed, which seamless integrating the KⅢnetwork into electronic nose system, making the electronic nose work more similar as a biological olfaction does.Furthermore, the pattern recognition method based on the KⅢnetwork is improved, through combining it with support vector machine to construct different series classifier and ensemble classifier. And their performances are explored in the case of classifying four kinds of teas and six different bands of fresh milk and constructed milk, respectively.更多还原