水作为生命之源与人类的生存息息相关, 近年来关于水环境污染的报道越来越多, 不容忽视。 实验以萘酚的两种同分异构体1-萘酚、 2-萘酚的混合物作为研究对象, 提出了一种新的算法, 通过对混合物的三维荧光光谱进行分析来实现水中萘酚的定性定量分析。 利用FS920稳态荧光光谱仪对配制的混合溶液进行扫描得到荧光光谱数据, 并对数据进行一系列的预处理去除拉曼散射和瑞利散射的影响。 将解决盲源分离(BSS)问题的独立成分分析(ICA)算法应用到荧光光谱定性定量分析问题当中, 盲源分离技术就是将测量得到的混合信号作为处理对象进行分解, 实现未知系统中源信号的求解, 并得到混合矩阵。 对混合物中单一物质的识别与测量与盲源分离问题类似。 采用基于负熵最大的快速独立成分分析(FastICA)算法对实验数据进行分解, 将所有样本的三维荧光光谱数据沿发射波长方向展开成为向量, 得到一个大小为( N× M)的矩阵( N为样本数, M为波长数), 将该矩阵作为快速独立成分分析的输入进行独立分量提取, 输出分别为单组分物质的展开荧光光谱和混合矩阵。 FastICA算法的关键是利用牛顿迭代算法得到解混矩阵, 但迭代过程中复杂的求导问题会使计算量增大、 迭代速度减慢, 针对该算法存在的问题, 提出用差分法(又称为双点弦截法)代替求导的解决方法。 为了验证算法的可行性, 用改进后的算法和原有算法分别对荧光光谱数据进行了五次独立分量提取实验, 原有算法平均运行时间为17.78 s, 而改进后的算法平均运行时间为3.22 s, 比原有算法提高了14.56 s, 有效地减少了计算量, 改善了FastICA算法的迭代速度并且使其收敛性更加稳定。 通过实验结果可以看出改进后的算法得到的光谱更接近真实的光谱。 利用快速独立成分分析算法分解得到的混合矩阵与物质浓度相关, 这是物质定量分析的依据, 但它们之间的关系可能是非线性的, 采用能实现非线性拟合的支持向量回归机(SVR)进行回归预测, 将混合矩阵和实际浓度矩阵分别作为SVR的输入和输出, 利用遗传算法(GA)对支持向量回归机的参数进行优化选择, 并选择径向基核函数(RBF函数)作为SVR的核函数, 建立回归模型, 实现对荧光光谱的定量分析。 1-萘酚的拟合相关系数( r)为0.998 6, 样品回收率(Recovery rate)为96.75%~104.2%, 预测均方根误差(RMSEP)为0.119 μg·L^-1; 2-萘酚的拟合相关系数为0.998 8, 样品回收率为96.8%~105.5%, 预测均方根误差为0.1 μg·L^-1, 预测结果比较令人满意, 符合预测要求。 实验证明改进的基于负熵最大的FastICA-SVR算法能实现对混合物中1-萘酚、 2-萘酚准确有效的识别和测量, 并且改进之后加快了算法的分解速度。

As the source of life, water is closely related to the survival of human beings. In recent years, there have been more and more reports on water pollution. Water pollution has become a serious problem, which can not be ignored. Two isomers of naphtol, 1-naphthol and 2-naphthol, were used as the research object in the experiment, and a new algorithm, which was used for qualitative and quantitative analysis of naphthol in water by analyzing the three-dimensional fluorescence spectrum of the mixture, was proposed. Using FS920 steady-state fluorescence spectrometer to scan the mixed solution and get the required experimental data. Then, a series of preprocessing steps for data are needed to remove the effects of Raman scattering and Rayleigh scattering. Independent component analysis (ICA) which is always used to solve the problem of blind source separation (BSS) will be applied to solve the problem in quantitative and qualitative analysis of fluorescence spectrum. BBS is an algorithm that uses the measured mixed signals as the processing objects to realize the decomposion of the source signals in the unknown system, as well as, to get the mixed matrix. The problem in identification and measurement of a single substance in a mixture is similar to the problem in blind source separation. The fast independent component analysis (FastICA) algorithm based on the maximum negative entropy is used to decompose the experimental data. The three-dimensional fluorescence data of all samples need to be expanded into a vector along the direction of the emission wavelength, and a matrix whose size is N× M can be obtained ( N is the number of samples and M is the number of wavelength). This matrix is used as the input of fast independent component analysis to extract independent component, and the output is the expansion fluorescence spectrum of the single component material and a mixed matrix. The key to the fast independent component analysis algorithm is using Newton iterative algorithm to obtain the solution matrix, but the complex derivation of iteration process makes this algorithm have some problems, such as large computation and slow iteration. In order to overcome the shortcomings of fast independent component analysis, the differential method, also called double point chord cut method, is proposed to replace the complex derivation problem in the iterative process. In order to verify the feasibility of the algorithm, five times independent component extraction experiments were carried out on the spectral data with the improved algorithm and five times independent component extraction experiments were carried out on the spectral data with the original algorithm. The average running time of original FastICA algorithm is 17.78 seconds, and improved FastICA algorithm is 3.22 seconds, which is 14.56 seconds lower than original algorithm. The experiment result proves that differential method instead of the complex derivation problem in the iterative process can effectively reduce the amount of calculation and improve the speed of the iteration of the fast independent component analysis algorithm and the convergence is more stable. It can be seen from the experiment result that the fluorescence spectrum which was obtained by the decomposition are closer to the real spectrum. The mixture matrix obtained by FastICA is related to concentration matrix, which is the basis for quantitative analysis of materials. But the relationship between the mixture matrix and the concentration matrix may be nonlinear. Therefore, it is necessary to take the nonlinear fitting method to realize the fitting between the two. Support vector regression (SVR) machine can realize nonlinear regression, so SVR will be used to obtain predicted concentration. The mixed matrix decomposed and the actual concentration matrix are as the input and output of support vector regression machine respectively. The parameters of SVR are crucial to the prediction. Genetic algorithm (GA) is used to optimize the parameters and radial basis function (RBF function) is selected as the kernel function of SVR. Then the regression model is established by using the algorithm to realize quantitative analysis of the fluorescence spectrum. The fitting correlation coefficient ( r) of 1-naphthol is 0.998 6 and 2-naphthol is 0.998 8; the recovery rate of 1-naphthol is 96.6%~104.2% and 2-naphthol is 96.8%~105.5%; the prediction of root mean square error (RMSEP) of 1-naphthol is 0.119 μg·L^-1 and 2-naphthol is 0.100 μg·L^-1. The results of the prediction are satisfactory and meet the requirements of the prediction. The experiment proved that the improved fast independent component analysis algorithm based on negative entropy combined with support vector regression algorithm can accurately identify and measure 1-naphthol and 2-naphthol in mixture, and this algorithm can also increase the speed of analysis for the hybrid system.