基于遥感FTIR技术的多组分污染气体浓度分布重构

【作者】 任翌博；

【导师】 王俊德； 李燕；

【作者基本信息】 南京理工大学 ， 应用化学， 2006， 硕士

【摘要】 遥感傅立叶变换红外光谱(Remote Sensing Fourier Transform Infared Spectroscopy，简称RS-FTIR)技术由于其独特的优势，越来越广泛地应用于室内、室外大气环境监测。本课题将人工神经网络建模与平滑基函数最小化重构算法(SBFM)结合，应用于RS-FTIR环境监测，实现同时重构多组分气体浓度在二维空间的分布。 本课题的研究主要包含以下几个方面： 1 单组分气体浓度分布重构 遥测遥感技术与计算机重构技术结合可以再现单组分气体浓度的二维空间分布。本课题研究在实验室内进行，首先，设计发散状实验光路，模拟污染气体发散实验，得到RS-FTIR仪扫描的吸光度谱图；然后，编写模拟退火优化的SBFM算法(SA-SBFM)程序，重构单组分气体浓度的二维空间分布。SBFM重构结果与RS-FTIR测量的光路积分浓度结果一致，证明了SBFM算法的可靠性；同时，通过比较多种重构模型的重构结果，得出双高斯重构模型最符合当前测量环境和测量要求。 2 RS-FTIR谱图定性分析 RS-FTIR与SA-SBFM结合，只能监测单种污染气体的浓度分布，为了同时定性、定量分析混合气体的红外遥感谱图，还需要借助有效的谱图解析方法。本论文将化学计量学方法中的两种神经网络建模方法——感知器神经网络和反传神经网络，应用于RS-FTIR谱图解析。 本课题研究建立感知器网络模型，将预处理后的光谱数据输入训练好的感知器网络，定性分析混合物中可能存在的三种气体——氯仿、甲醇、二氯甲烷，定性分析的准确率达到100％。 为了定性分析重叠更严重的五组分FTIR谱，主成分分析方法(Principle Component Analysis，简称PCA)被用来提取光谱数据主成分，作为感知器网络的输入，建立混合气体FTIR吸光度的一阶导数与混合气体成分的线型模型，定性分析准确率同样100％。PCA的应用大大减少了运算时间，且网络泛化能力增强，可以准确预测出训练集没有的混合气体类型。 3 RS-FTIR谱图定量分析 在定性分析的基础上，我们选用反向传播神经网络(Back Propagation-Artificial Neural Network，简称BP-ANN)定量分析三组分混合气体(氯仿、甲醇、二氯甲烷)各组分的含量。由于甲醇的特征吸收波长与其它两种气体无混叠，我们同时用Lamber-Beer定律定量计算组分含量，并与ANN的预测结果进行比较，两种分析方法结果的一致性验证了神经网络是一种有效的多组分定量分析方法。更多还原

【Abstract】 Remote Sensing Fourier Transform Infrared Spectrometry (RS-FTIR), for its unique advantages, has found increasing application for indoor, outdoor air monitoring. In this research, two strategies of Artificial Neural Network (ANN) models with Smooth Basis Function Minimization (SBFM) reconstruction algorithm were combined in RS-FTIR spectra analysis, for the purpose of monitoring multi-components Volatile Organic Compounds (VOCs) concentration distribution in the two-dimensional plane. The main work is shown below:1 Reconstruction of single-component VOC’s concentration distributionPrevious studies showed that Optical Remote Sensing and Computed Tomography could be connected to measure the spatial distribution of single-component VOC concentration in atmosphere. In this research, chamber experiments were conducted to test the combination of mono-static RS-FTIR spectroscopy and SBFM algorithm with radial beam path design. Through chamber experiment, VOC’s absorbance spectra were obtained, and input to SBFM reconstruction program with the optimizing algorithm of Simulated Annealing (SA). The reconstruction results showed excellent agreement with the ray integral concentrations measured by RS-FTIR. Meanwhile, three models’ reconstruction results were compared, and the Double-Gaussian reconstruction model was found to fit the best to the current monitoring environment.2 Qualitative analysis of RS-FTIR spectraWith the combination of RS-FTIR and SA-SBFM, we could just monitor single gas contaminant’s concentration distribution. For identifying and quantifying multi-component VOCs from the complicated mixture RS-FTIR spectra, we still needed the assistance of advanced spectra analyzing methods. In the last decade, the number of chemometric methods applied in the field of spectroscopy grew rapidly. In this research, we tested two ANN modeling strategies—perception neural network, and Back Propagation network, in use of RS-FTIR spectra analysis.Perception network was designed to identify each component from three-component VOCs mixture spectra containing chloroform, methanol and methylene chloride. The responses of FTIR spectra data were normalized and taken first derivation, then input to supervised training network. Its identifying accuracy is 100%.For analyzing more complicated and overlapped 6-component spectra containing Benzene Toluene Chlorobenzene, 1,2-diethylbenzene and Cyclohexane, Principal更多还原