【作者】 王鹤；

【导师】 周东祥；

【作者基本信息】 华中科技大学 ， 微电子学与固体电子学， 2007， 博士

【摘要】 电子鼻是一种由具有部分选择性的气体传感器阵列和适当的模式识别技术组成,能够识别简单或复杂气体的仪器。它在环境监测、食品工业、化学工业等领域有广泛的应用前景。因此,这一研究领域受到广泛关注和普遍重视,近年来发展迅速。作为检测气体的电子仪器,电子鼻是气体传感器阵列和信息处理技术的有效结合,因此对模式识别技术的研究显得尤为重要。本论文从生物嗅觉系统的结构和嗅觉机理出发,研究了电子鼻的原理和结构,探讨了传感器阵列的组成,特别是在模式识别方法的选择和改进上,对传统的BP神经网络进行优化改进,提出了基于遗传神经网络的电子鼻系统方案,以提高混合气体检测的速度和精度。主要进行了以下四大方面的研究:(1)研制了一套电子鼻测试系统。主要包括气体传感器阵列、配气系统、气体检测系统硬件及气体检测系统软件。在分析了气体传感器的检测原理和构成阵列的准则的基础上,构建了气体传感器阵列;设计了配制混合气体(H2S, CO, CH4 )的配气系统(所测的气体浓度范围是H2S:0-100ppm,CO:0-100ppm,CH4:0-500ppm);自行配置的数据采集装置,为混合气体的检测打下了基础。(2)模式识别方法的选择。深入分析了BP神经网络的基本工作原理,结合本电子鼻系统的特点,将BP神经网络应用于混合气体的检测中,并进行了定性分析和定量分析。(3)提出了一种遗传算法优化神经网络(GANN)的气体检测方案。针对BP算法的不足,提出了遗传算法改进神经网络的方案,分两步对神经网络进行权值优化,第一步用遗传算法内嵌神经网络,搜索出神经网络权值大致范围内的最优个体;第二步以遗传算法搜索出的最优个体作为神经网络的初始权值,训练网络。实验证明,采用此气体检测方案,可以充分结合遗传算法的全局搜索能力和BP算法的局部搜索能力,加快收敛速度(迭代次数从144降到52),提高气体检测精度(三种气体平均误差分别从6.54ppm、6.82ppm和28.83ppm降到4.64ppm、4.37ppm和17.13ppm)。(4)设计了一种改进的遗传算法自适应变异(IAGA)气体检测算法。在理论分析遗传算法各参数对性能的影响基础上,提出改进自适应变异算法,通过计算进化效率,自适应调整变异率和变异量。只需考虑进化代数这一个因素,而不必考虑影响遗传算法性能的诸多复杂因素。计算量小、通用性强且效率高。实验结果表明,该气体检测方法可以进一步增强神经网络对训练样本的收敛性(收敛成功率从40%提高到75%)和对测试样本检测的准确性(三种气体平均误差分别降为3.72ppm、4.22ppm和15.78ppm),提高了气体检测的效果。更多还原

【Abstract】 Electronic Nose (EN) is an instrument, which comprises an array of gas sensors with partial specificity and an appropriate pattern recognition system, capable of recognizing simple or complex odours and gases. EN technology is a cross field with the technologies of sensors, electronic technique, computer, pattern recognition. EN is applied wide in the fields of environmental detection, food and chemistry industry etc with its unique function. What’s more, it is getting attention with more domains.As an electronic instrument of detecting gas, EN is the combination of gas sensor array and information processing; therefore the research of pattern recognition technology is of great value. The EN is researched roundly in this paper from the four aspects: the basic principle of EN, the implementation program, the choice and improvement of pattern recognition technology in gas detection.The principle, implementation and design though of the EN were generalized overall from the mechanism of biologic olfactory system. After analyzing the pattern recognition methods, the artificial neural network and genetic algorithm was chosen as the gas detection method of the EN system.A set of EN detection system was designed and realized including the gas sensors array, gas distribution device, test hardware and software. The detection principle of gas sensors and guideline of composing sensor array were analyzed emphatically; the gas distribution device was ready for the preparation of mix-gas(H2S, CO, CH4), the concentration ranges of the three gases are 0-100ppm, 0-100ppm, 0-500ppm respectively. The software of data acquisition is the foundation of mix-gas detection.As for the research of the pattern recognition, the neural network based on the back propagation (BP) neural network was applied in the mix-gas detection, the qualitative and quantitative analysis showed that the neural network can used in the mix-gas detection. The shortage of BP algorithm and the revised measure were analyzed. The BP algorithm is the local search method based on the gradient descent, to overcome the local search character,the genetic algorithm was applied to optimize the BP neural network.The genetic algorithm optimizing neural network was put forward to optimize the neural network weigh coefficient with two steps. The first step was to embed the neural network with genetic algorithm, searching the optimum individual in the approximate range of neural network weigh coefficient; the second step was to use the optimum individual as the initial weigh value, then training the network. The experiment shows that the genetic algorithm optimizing neural network can combine with the global search character of genetic algorithm and local search character of BP algorithm. The genetic neural network can accelerate the speed of convergence (the iterative number descended form 144 to 52) and improve the accuracy (the average error of three gases descended from 6.54ppm,6.82ppm,28.83ppm to 4.64ppm,4.37ppm,17.13ppm respectively).An improved self-adaptive mutation of genetic algorithm was designed with the theory analysis of the parameters of genetic algorithm affecting the work performance. The self-adaptive mutation operate has the ability of jumping over the local minimum point and maintaining the diversity of population to realize the global search. Accordingly the modified self-adaptive mutation operate was put forward. It was a self-adaptive algorithm based on evolution efficiency, during computing the efficiency, the mutation ratio and mutation quantity can self adapt. The algorithm can be only thought about the evolution generations, which is simple and easy to manipulate. Applying the modified algorithm into the mix-gas detection shows that the network accelerates the convergence performence (the success rate increased from 40% to 75%) and improves the detection accuracy (the average error of three gases descended to 3.72ppm, 4.22ppm, 15.78ppm respectively).更多还原