【作者】 胡兰萍；

【导师】 王俊德；

【作者基本信息】 南京理工大学 ， 化学工程与技术， 2007， 博士

【摘要】 本文利用化学计量学的方法，结合遥感傅里叶红外光谱(Fourier transform infrared Spectrum，FTIR)技术，对大气环境中的有毒易挥发有机化合物(Volatile Organic Compounds，VOCs)进行了较为深入的定性、定量研究。本文的主要研究内容如下：1．基于化学计量学的多组分遥感FTIR谱图解析针对人工神经网络(ANN)训练时间过长和模型“过拟合”的问题，采用偏最小二乘法(PLS)和主成分分析法(PCA)，对输入ANN的光谱数据进行了主成分提取，利用遗传算法(GA)进行波长选择，以简化模型、剔除不相关的变量、加快分析速度。建立了PLS-BP-ANN、PCA-BP-ANN、GA-BP-ANN和BP-ANN四种模型，对谱带混叠严重的多组分体系成功地进行了同时定量分析。在苯、甲苯、甲醇、氯仿和丙酮五组分大气VOCs的定量测定中，对四种方法的预测误差比较，PLS-BP-ANN模型显示了最为稳健的能力。2．模型传递方法用于解析遥感FTIR谱图本文将模型传递的思想引入到遥感FTIR谱图的分析中，模型传递就是将一台仪器所获得的数据建立的校正模型，直接用到另一台仪器测得的数据上，旨在简化分析过程并获得好的预测精度，这样就可消除同一样品在不同仪器上的测量误差。而遥感FTIR的问题，本文将其理解为样品在实测中相对于标准红外谱图所出现的偏差。采用普鲁克分析(PA)就是去除仪器间的差异，PA的基本思想是去除X(如吸光度)中与Y(如浓度)不相关的部分。由于其更针对在不同仪器上的量测信号进行处理，消除这些信号间的差别，同时具有较好的适应性，因此是一种更有意义的模型传递方法。本研究利用PA实现了四组分—丙酮、苯、氯仿和甲醇气体混合物的BP-ANN模型在两台FTIR仪器上的传递，预测精度高，实现了EPA数据对遥感数据的分析。3．人工神经网络研究大气VOCs的分类针对BP-ANN学习收敛速度慢、建模过程中网络层数及每层神经元的个数难以确定和易陷入局部最优等问题，本文提出了适用于模式分类的径向基神经网络，即概率神经网络(Probabilistic Neural Network，PNN)和学习矢量量化(Learning Vector Quantization，LVQ)神经网络的方法，对大气环境质量进行预测。在氯仿、甲醇、丙酮、己烷、甲苯和二氯甲烷六组分大气VOCs的定性测定中，对PNN、LVQ和BP-ANN三种方法的预测误差进行比较，结果表明：PNN方法的分类精度最好，达到93.3％。而且由于PNN的结构简单、训练速度快、新的训练样本也很容易加入到以前训练好的分类器中，因此它很适合于大气环境的实时、在线定性监测，能起到及时预警的作用。4．遥感FTIR谱图的模式识别本文创造性地将主成分提取与线性判别分析(Linear Discriminant Analysis，LDA)相结合的方法，引入到遥感FTIR光谱的分析中，建立了PLS-LDA、PCA-LDA对VOCs的模式识别方法。在对谱图相互严重混叠干扰的五组分—己烷、苯、甲苯、丙酮和二氯甲烷大气VOCs的定性鉴别中，两种方法的识别率都较高，其中PLS-LDA略高于PCA-LDA，该项研究对工作和生活空间的污染控制具有指导意义。本文将这种方法推广到其它复杂体系的定性分析中，如对多环芳烃(Polycyclic Aromatic Hydrocarbons，PAHs)的致癌性进行分类，致癌性按高(h)、低(l)、非(n)分类时预测准确率达100％。更多还原

【Abstract】 Qualitative and quantitative analysis of VOCs (Volatile Organic Compounds)in the atmosphere are carried out in combination with remote sensing FTIR(Fourier transform infrared spectrum) technology and chemometrics methods inthis dissertation. The main contents are as follows:1. Interpretation on Remote Sensing FTIR Spectrum Based on ChemometricsIn consideration of ANN’s weakness of excess training time and overfittingmodels, this dissertation adopts PLS and PCA to extract principal component fromANN spectrum data, and uses GA to choose wavelength to simplify models, toeliminate irrelevant variables, and to enhance analytic speed. Four models,PLS-BP-ANN, PCA-BP-ANN, GA-BP-ANN and BP-ANN, are built to performsimultaneous quantitative analysis of multicomponents in the mixture when thereare serious overlapping between the spectral bands of the compounds Comparingthe predictive error of the four methods with the quantitative measurement for thefive-component atmosphere VOCs(benzene, toluene, methanol, chloroform andacetone), PLS-BP-ANN model had the best robustness.2. Calibration Transfer for the Interpretation on Remote Sensing FTIRSpectrumThe idea of calibration transfer is introduced into the analysis of remotesensing FTIR spectrum this dissertation. Calibration transfer means to predictsuccessfully the signals of other uniform instrument with the calibrated modelbuilt with one instrument. As for the data of remote sensing FTIR, it is regarded asthe deviation from EPA data in this research. And then the method of PA(Procrustes Analysis) is adopted to get rid of discrepancy among instrumentswhile its main idea is to eliminate the parts in X (such as absorbance) irrelevant toY (such as concentration). PA focuses on the process of metrical signals ofdifferent apparatus to get rid of discrepancy among them and maintains bettersteadiness. Hence, it is a more significant method of calibration transfer. Thisresearch adopts the method of PA to realize the transfer of BP-ANN model madeup of acetone, benzene, chloroform and methanol between two FTIR apparatus. Ityields high prediction accuracy and fulfills the prediction of remote sensing data with EPA data.3. Classification of VOCs in the Atmosphere Based on Artificial NeuralNetworkConsidering the confinements of slow speed of study convergence, too muchlayers of network and local optimization with BP-ANN model, the improved ANNmodels, PNN(probabilistic neural network) and LVQ(learning vector quantization)which are suitable to pattern classification are built in this dissertation. For thesix-component system chloroform, methanol, acetone, hexane, toluene andmethyene chloride in the atmosphere, the qualitative measurement performance ofPNN, LVQ, and BP-ANN are compared while the method of PNN maintains thebest classification accuracy with 93.3%. Besides, PNN is characterized of simplestructure and fast training speed. And the new training samples are easier to addinto the former trained classifiers for PNN model. Consequently, PNN is suitablefor the real-time and on-line qualitative monitoring of atmospheric environmentand can perform environmental alarm.4. Pattern Recognition of Remote Sensing FTIR SpectrumThe method of PCA-LDA (principal components analysis-linear discriminantanalysis) is introduced creatively in the analysis of remote sensing FTIR spectrumin this dissertation and the method of pattern recognition of PLS-LDA andPCA-LDA are built to recognize VOCs in the atmosphere. In the qualitativedifferentiation of five components with overlapped spectrum, that is, hexane,benzene, toluene, acetone and methyene chloride, both methods have high ratio ofrecognition and PLS-LDA yields a little higher one than PCA-LDA. This researchis of great significance to the control of pollution in our life. This method is alsopromoted to qualitative analysis of other complex system in this research such asthe classification of carcinogenicity of PAHs (polycyclic aromatic hydrocarbons).When the carcinogenicity is classified on high, low, and no level, the predictiveaccuracy reaches 100%.更多还原