基于BP神经网络的VOCs混合气体检测研究

发布时间：2018-12-13 13:32

【摘要】：为实现对挥发性有机化合物(VOC)混合气体检测,本研究采用传感器阵列和模式识别技术相结合的电子鼻方式对问题进行探究。传感器阵列是在实验室中自制的旁热式金属氧化物半导体传感器组成的阵列,能对VOC混合气体形成完备的响应模式。传感器阵列在不同VOC混合气体响应数据集均来源于实际实验测试。实验中为其问题的实验探究,自行搭建了传感器测试系统,其中的VOC混合气体由乙醇,丙酮,甲醛和甲苯四种典型VOC气体组成的目标分析物,并且为了开发用于实际应用的电子鼻,每种VOC及其组合的浓度在混合物中随机分配。探究模式识别时,本文采取BP神经网络对传感器阵列信号进行分析和识别,完成对VOC混合气体组分及浓度估计。BP神经网络在MATLAB中分析创建,首先须要做的是对数据归一化预处理,以防止因数量级别而造成计量误差,然后还探究了BP神经网络中隐含层神经元个数,激活函数,性能目标等结构参数对网络预测性能的影响,调试出适合本问题的最佳结构参数。根据实验结果可知,BP神经网络输出节点对目标分析物中每种VOC的浓度可以给出连续预测,且在一定误差范围内,它是能够完成对VOC混合气体组分的定量分析。本文以提高系统预测精度为目的,对模式识别中方式进行改进,思虑先由决策树以VOC总量为标准对VOC混合物数据集进行等级分类完成初步筛选,再由BP神经网络基于不同等级,调试出适当的结构参数,并对等级内的样本针对性训练,完成浓度估计。实验结果显示,改进后的模式在每种VOC的浓度估计中的最大误差约为2ppm,其精度优于从单个BP神经网络获得的结果。此外,当预测的浓度高于20ppm时,相对误差小于5%。这项研究表明:神经网络定量分析VOC混合物浓度方面的潜力。改进后的模式能较准确地完成VOC混合气体组分的定量分析,成为开发针对VOC气体进行识别的电子鼻产品的基础。
[Abstract]:In order to detect the mixed gases of volatile organic compounds (VOC), an electronic nose combined with sensor array and pattern recognition was used to study the problem. Sensor array is an array of side heat metal oxide semiconductor sensors made by ourselves in laboratory. It can form a complete response mode to VOC mixed gas. Sensor arrays in different VOC gas mixture response data sets are derived from the actual experimental test. In order to explore the problem in the experiment, a sensor testing system was set up, in which the VOC mixture gas was composed of four typical VOC gases, ethanol, acetone, formaldehyde and toluene. In order to develop an electronic nose for practical applications, the concentrations of each VOC and its combinations are randomly distributed in the mixture. In this paper, BP neural network is used to analyze and recognize the sensor array signals, and the mixture gas components and concentrations of VOC are estimated. BP neural network is established in MATLAB. The first thing we need to do is to preprocess the data normalization so as to prevent the metrological error caused by the quantity level. Then we also explore the number of neurons in the hidden layer and the activation function in the BP neural network. The effect of performance target and other structural parameters on the network prediction performance is studied and the optimal structure parameters suitable for this problem are debugged. According to the experimental results, the output node of BP neural network can give a continuous prediction of the concentration of each VOC in the target analyte, and within a certain error range, it can accomplish the quantitative analysis of the VOC mixture gas component. In order to improve the prediction accuracy of the system, the method of pattern recognition is improved in this paper. First, the decision tree classifies the VOC mixture data set according to the total amount of VOC, and then the BP neural network is based on different grades. The appropriate structural parameters were debugged and the samples in the grade were trained to estimate the concentration. The experimental results show that the maximum error of the improved model in each VOC concentration estimation is about 2 ppm.The accuracy of the improved model is better than that obtained from a single BP neural network. In addition, when the predicted concentration is higher than 20ppm, the relative error is less than 5. This study shows the potential of neural networks for quantitative analysis of VOC mixture concentrations. The improved model can accurately accomplish the quantitative analysis of VOC mixture gas components, which is the basis of developing electronic nose products for VOC gas recognition.
【学位授予单位】：宁波大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP212.9;TP183

【参考文献】