| With the rapid development of modern analytical technology, analysts can now easily and quickly obtain more data points, and the data sets is becoming complex more and more. These data arrays not only contain amass of chemical information, but also involve some interferers including irrelevance chemical components, noise and background. The traditional methods of the processing data for analytical method can not meet the analysis needs. The use of mathematical seperation instead of "chemical or phisical seperation" for quantification resolution of overlapped spectra of complex matrices could make multi-component analysis with the characteristics of being simple, direct and faster. The curve fitting and radial rasis function neural network are used for molecular fluorescence analysis. It provides a new way and method for the application of spectroscopy.First of all, it is described that the history and development of chemometrics, and the research advance of multivariate resolution and calibration in spectrum are reviewed. The basic principle of the related algorithms is introduced in the second chapter.In Chapter3, the radial basis function artificial neural network based on genetic algorithm was combined with capillary electrophoresis for quantitative resolution of overlapped peaks in electrophoretogram. It is achieved that the simultaneous and accurate quantification of hardly separating dihydroxhenzene, ohenol and p-nitrophenol.Chapters4and5used the exponentially modified Gaussian (EMG) model-based genetic algorithm as a fitness function for fitting fluorescence spectrogram. The method was effective for solving the interference of fluorescent substance in the course of the multi-component quantitative analysis. As a example, the interference of endogenous fluorophores in different urines on fluorescence of gatifloxacin (GFLX) was examined by using the fitting fluorescence spectrogram. Another example, the proposed method can effectively correct the overlapping interferences of fluorescence spectra of the three isomers. Under the optimized experimental conditions, the good linear relationship between the fluorescence intensity and concentration of GFLX concentration was obtained in the range of0.06μg·mL-1-3.5μg·mL-1with a correlation coefficient of0.9994. The detection limit and recovery were0.02μg·mL-1and99.2%-109.4%, respectively, with the relative standard deviation from1.3%to2.7%. In addition, the good linear relationship between the fluorescence intensity and concentration of catechol, resorcinol and hydroquinone was obtained in the range of0.02μg·mL-1-10μg·mL-1,0.01μg·mL-1-10μg·mL-1and0.01μg·mL-1-10μg·mL-1with a correlation coefficient of0.9920,0.9990and0.9996, respectively. Their detection limits were0.005、0.003and0.002μg·mL-1, respectively, and the recoveries were in the range of84.0%-117%with the relative standard deviation of0.3%-2.9%. The proposed fitting fluorescence spectrometric method was rapid, simple and highly sensitive for the determination of GFLX in different human urine and catechol, resorcinol and hydroquinone in water without preseparation. The results are satisfactory.In Chapter6. the effects of noise level and resolution between sub-peaks on modeling of the General Regression Neural Network (GRNN) and Back Propogation (BP) artificial neural network were investigated in detail by simulated data based on involved references. The different noise and resolution of spectrums were simulated by computer. The data points corresponding spectrum selected evenly were used as input variables of the neural network, and neural network model was trained. Prediction error of neural network was compared and the relation of the recognition capability and noise and total resolution of overlapped spectroscopy was discussed to provide valuable reference for application of the neural network. It was shown that under the training conditions provided in this work the recognition capability of the two artificial neural networks reduced gradually with increasing noise level, and the accuracy of the quantification results by General Regression Neural Network model was improved with the increase of the total resolution.In Chapter7, the genetic algorithm (GA) was used in optimizing input variables of radial basis function (RBF) artificial neural network to improve the precision of quantitative analysis of the unresolved spectra by artificial neural network. Using the proposed method can, in some extent, increase prediction capability, reduce "over-fitting" of the trained networks and structure of RBF artificial neural network, and improve learning ability of artificial neural network. Effective quantification analysis of overlaped synchronous fluorescence spectrms was achieved. This work provides a basis on the combination of theory and experiment for the application of RBF artificial neural network in spectroscopy.In Chapter8, the principal components analysis (PCA) was used in optimizing input variables of the radial basis function artificial neural network to improve the recognition capability. The irrelevant data points were removed by PCA. Therefore learning ability of artificial neural network can be improved to some extent. Use of the presented method can eliminate effectively the interference of endogenous fluorophores in urines on fluorescence of Norfloxacin (NOR). A new fluorescence method for the determination of NOR in urines was developed. Under the optimized conditions, the prediction error of Neural Network model for NOR was15.32%, and Neural Network Structure was2:3:1。 The method is quick and convenient for the determination of Norfloxacin in urine without interference。In Chapter9, the two kinds of data compression technology (PCA and GA) are widely used for feature extraction. The GA and PCA were used in optimizing same input variables of the radial basis function artificial neural network, respectively, to improve the recognition capability of neural network. The same simulated and experimental data were optimized by GA and PCA, and neural network models were trained. The prediction error of the two neural network models for three isomers and Neural Network Structure were calculated and compared. Under the optimized conditions, after obtimized with PCA, the prediction error is16.1%(simulated data) and17.81%(experimental data), and the structure of neural network model is7:14:3(simulated data) and8:22:3(experimental data). The results indicate that the use of PCA for the optimization of neural network has better neural network structure than use of GA, and the use of GA for the optimization of neural network has higher recognition capability than use of PCA. |
PDF Full Text Request