【作者】 彭姝迪；

【导师】 陈伟根；

【作者基本信息】 重庆大学 ， 电气工程， 2010， 硕士

【摘要】 随着科学技术的进步和工业生产的发展,对气体的检测不仅要求快速、准确,而且还要求检测能够在线进行。由于气体传感器普遍存在交叉敏感,单一气体传感器无法对多种气体进行准确的定性识别和定量检测。因此,将多个气体传感器组成阵列与模式识别方法相结合进行多组分气体的识别与检测具有更大的实用价值。而模式识别技术对识别效果起着关键性的作用。本文以电力变压器油中六种溶解气体(H2,CO,CH4,C2H4,C2H2,C2H6)在线检测为例,提出了一种基于免疫加权支持向量回归机的气体传感阵列模式识别方法。论文依次讲述了气体传感阵列国内外研究现状和常用模式识别方法,并对MQ型气体传感器进行特性实验研究,介绍了支持向量机模式识别理论,最后着重对基于支持向量机的气体传感阵列模式识别方法进行了逐步深入的研究,在利用气体传感阵列获取气体组分和浓度多维信息的基础上,结合支持向量机模式识别理论,对变压器油中溶解气体进行定量分析。首先对MQ型气体传感器进行了特性实验,通过实验验证了MQ系列气体传感器具有良好的灵敏度、重复性和响应特性,但同时存在一定的交叉敏感。针对神经网络模式识别方法存在网络结构难于确定、易陷入局部极小值等问题,分别采用BP神经网络和标准支持向量回归机(SVR)对气体传感阵列测得的信号进行模式识别。结果表明,经过SVR进行模式识别后的值明显优于BP神经网络识别值。将免疫算法用于支持向量机参数优化过程中,利用免疫算法保持群体多样性的特点进行参数全局寻优,将参数优化后的支持向量机应用于气体传感阵列模式识别中。通过对比表明,免疫支持向量机克服了标准支持向量机(SVM)采用试凑法确定参数的不足,识别精度较SVM有进一步的提高。针对标准支持向量回归机中未考虑各样本重要性的差异问题,给各个样本的惩罚系数赋予不同权重,分别采用线性插值法和非线性插值法对参数进行加权,将经过免疫算法全局寻优和参数加权后的支持向量机应用于气体传感阵列模式识别中。结果表明,免疫加权支持向量机具有更高精度的识别效果,更好的性能和应用前景。更多还原

【Abstract】 As the advancement of science and technology and the development of industry, the detection of gases requires not only fast, accurate, but also testing on-line. Caused by the cross-sensitivity of the gas sensors, it is impossible for a single gas sensor to identify multiple gases. So, identification of multi-component gas based on gas sensor array and pattern recognition is becoming an important way in dealing with cross-sensitivity in gas analysis. The pattern recognition technology plays the crucial role to the recognition effect.Taking the detection of six kinds of transformer oil dissolved gas (H2, CO, CH4, C2H4, C2H2, C2H6) as an example, this article proposed a pattern recognition method of gas sensor array based on immune weighted support vector machine for regression. The current research situation and commonly used pattern recognition methods are described at the beginning of the thesis. Characteristic experiments of MQ type gas sensors are carried. The author introduced the theory of support vector machine and pattern recognition method of gas sensor array based on SVM is studied step by step. The theory of SVM is used to do quantitative analysis after obtaining the multi-dimensional information of the gas composition and concentration by gas sensor array.Firstly, characteristic experiments of MQ type gas sensors are carried. Research shows that these gas sensors have good performance on sensitivity, repeatability and response characteristic. But cross sensitivity also exists at the same time.Secondly, aiming at the problems such as difficult determination of net structure and local minimization of neural networks, the BP neural networks and the standard SVR are used in the gas sensor array signals pattern recognition. The result shows the values of standard SVR is better than BP neural networks. Immune algorithm is used in the SVM parameters optimization process. Because of immune algorithm can maintain the diversity of the group, the parameters can be adjusted globally and use it in sensor array signals pattern recognition. Results show that, this kind of pattern recognition method overcomes the shortages of the trial-and-error method, and improves recognition accuracy of SVM.Finally, aiming at the problem of no considering the importance of each sample in the standard SVR, each training sample is assigned different approximation error requirement and different penalty. the thesis take linear interpolation method and nonlinear interpolation method to the weighted parameters and use the SVM which combine immune algorithm with parameters weighted in the gas sensor array signals pattern recognition. The results show that, immune weighted support vector machine have better recognition accuracy, performance and application prospect.更多还原