【作者】 张琳；

【导师】 王俊德；

【作者基本信息】 南京理工大学 ， 化学工程与技术， 2006， 博士

【摘要】 本论文用化学计量学的方法，建立了对大气环境中有毒易挥发有机化合物(volatile organic compounds，VOCs)的遥感傅立叶变换红外光谱(Fourier transform infrared，FTIR)谱图解析方法，文中所建立的方法，更好地发挥了遥感分析的快速、准确、实时等优势。同时，化学计量学方法在具体问题的解决中，也得到发展。由于遥感FTIR谱图存在信号噪声大、未知干扰、背景信号波动大等特点，本文从方法的普适性和稳健性角度，提出了遥感FTIR谱图的解析的方法。建立了基于偏最小二乘法(partial least squares，PLS)的谱图定量分析方法：正交信号处理(orthogonal signal correction，OSC)方法较好地对干扰和信号进行分离，得到了一种稳健和简单化的模型；由改进PLS线性内部关系出发，建立了对5组分VOCs体系进行分析的多项式偏最小二乘(polynomial partial least squares，PPLS)的方法；创新性地引入模型传递的思路，通过选择方法以及方法的优化，用少的潜变量建立分析模型并实现了EPA数据对遥感数据的分析；用PLS方法建立对VOCs的模式识别方法，结合实验设计手段实现对VOCs定性和定量同时分析；本研究初步建立了被动式遥感FTIR谱图分析技术，创新性地利用平行因子分析法(parallel factor analysis，PARAFAC)实现了在有干扰情况下，被动式遥感FTIR谱图的定性和定量的同时分析。本文取得的基本成果总结如下： 1．基于偏最小二乘法的遥感FTIR谱图解析方法 本研究根据遥感FTIR信号特点，从方法的普适性和稳健性角度出发，建立了改进PLS的谱图分析技术。对于遥感FTIR数据，由于噪声的加入，使得PLS模型的潜变量变大，噪声数据不能同真实数据相分离，OSC-PLS有效改善了对遥感FTIR谱图的解析能力，建模潜变量个数的减少，对更复杂的体系也同样具有较好的预测性能。相对于PLS，PPLS利用非线性的内部关系，对VOCs混合物含量的预测准确度有了提高，显示出很好的处理非线性数据的能力。遗传算法(genetic algorithm，GA)可以有效选取PLS建模变量，使模型简单化的同时提高预测准确度。但是，GA-PLS对遥感数据的预测仅限于用遥感数据建立校正模型的情况。经过PLS对数据压缩，人工神经网络(artificial neural network，ANN)输入数据变小，计算精度提高。既改善了ANN模型对训练集的要求，又满足了遥感FTIR实时分析的要求。更多还原

【Abstract】 Techniques were set up in the dissertation with the aid of chemometrics, which were used for interpretation on remote sensing FTIR spectrum of VOCs (volatile organic compounds) in the atmosphere. These especially strengthen the advantage of remote sensing FTIR in quick, accurate, real-time and simultaneous determination of multi-component analysis. Since the remote sensing FTIR spectrum containing much noise, unknown interferents and background shift, strategies were put forward in the viewpoint of model robust and popularity. Based on the multivariate calibration method PLS (partial least squares), new signal correction method OSC (orthogonal signal correction) was applied to the filtration of noise or other unrelated parts and good performance was yielded in model simplicity and robust. Furthermore, PLS were improved in the direction of inner relation with the use of PPLS (polynomial partial least squares) model and were employed in the five-component mixture successfully. Calibration transfer was applied innovatively in the analysis of the remote sensing FTIR spectrum. With the optimization of this method, remote sensing FTIR spectra were analyzed directly with EPA spectra. PLS was also used for the pattern recognition of VOCs, and qualitative and quantitative analysis were achieved simultaneously with the help of experiment design. Primary interpretation of passive remote sensing FTIR spectrum was carried out. The innovation-application of PARAFAC (parallel factor analysis) was utilized for the analysis of passive remote sensing FTIR spectrum and qualitative and quantitative analyses were realized. The main conclusions were achieved as following:1. Interpretation on Remote Sensing FTIR Spectrum Based on PLSAccording to the signal characteristics of remote sensing FTIR, PLS was improved in the viewpoint of model popularity and robust. As for the remote sensing FTIR data, noise in the data enlarged the LVs (latent variables) of PLS and made it difficult for the separation of noise and information. OSC-PLS modified this situation in prediction accuracy and fewer latent variables and efficient filtration, which was validated by the complex system. Compared with PLS, the inner relation of PPLS was nonlinear, which was used successfully for the prediction of five-component system. Variables of PLS was optimized by GA (genetic algorithm) and then made the model simple and accuracy. However, this method was only suitable for the prediction of remote sensing data with remote sensing data itself. With the data compressed by PLS, the size of ANN (artificial neural network) input was smaller and prediction accuracy更多还原