Study And Application Of Data Processing Techniques For SERS Spectrum

Raman spectroscopy (RS), a technique with advantages of label-free, non-invasive, high molecular selective and insensitive to water, is capable of providing molecular fingerprint information of biological systems with high spatial and temporal resolution. RS has become a powerful potential method in fast detection. The Raman intensity can be increased by106-1015with the substrates of roughened metal (gold, silver or copper) electrodes, which is the surface enhanced Raman spectroscopy (SERS). Nowadays SERS has shown its huge potential for the detection of trace amount in fields of food safety, environment monitoring, surface science, material analysis, biology and medical science, etc. However, the spectrum recognition is challengeable and tough because of several factors including the low reproducibility of the enhancement, the disturbance of coexisting substances in complex matrices. Several key signal processing and recognition techniques were studied in this thesis, and they have been used in the software of SERS spectrum analysis.Below is the main contents of this thesis.1. The preprocessing methods were explored in order to decrease the influence of fluorescence background on the final analysis results. A series of preprocessing algorithms are proposed including window width adapted polynomial moving average filtering, robust baseline correction method based on iterated Fourier transform, baseline correction method based on1-order derivative smoothing skill. With small adjusting of several parameters, the algorithms show good results when used in various detection matrices. Furthermore, it is shown that the two baseline correction methods apply to spectra with different baseline trends.2. In terms of the feature extraction for the Raman spectrum, several techniques are discussed.1. A distance discriminant method based on the within-class and between-class distances is put forward. It is used to explore the best principal components assemble with minimum redundancies when using the principal component analysis (PCA) method.2. The independent component analysis (ICA) method is used for the matrices where the character peaks of solvent and detected substance are overlapped. In these cases the overlapped Raman signals can be separated after ICA. The algorithms are used for the detection of Thiocyanate Ion in milk and proved to successfully improve the qualitative and quantitative analysis accuracy.3. A peaks fitting method combined with intensity normalization skills is presented for the cases where character peaks of detected substance are very few and the Raman signals hardly vary with density. It performs well when used for the detection of Arsenic in drinking water.3. Classifiers based on techniques of the back-propagation artificial neural network (BP-ANN), the support vector machine (SVM) and the linear discriminant analysis (LDA) are researched and modified. For the BP network, a two-step training method is studied to solve the problems of local minimum and slow convergence. While a global error based weight adjusting way is used in order to avoid the effects of processing order of the training samples. Lastly, the PCA-LDA technique is used to solve the singular difficulty of small sample data problems. This technique is proved to be able to achieve the qualitative analysis of the hard separated matrices. However, all of the results we have obtained tell us different classifiers are needed in different matrices in order to get the best classified results.4. Considering the low reproducibility of SERS, a first step is taken for the application of signal processing techniques in the quantitative application of SERS. The ICA and linear calibration based method is proposed for the quantitative analysis of Thiocyanate Ion in milk. The accuracy is finally improved.5. The algorithms in this thesis are used in the SERS instrument for the detection of Malachite green in seawater, Rhodamine B in pepper and Arsenic in drinking water with the accuracy above97%. It is proved to meet the requirement of fast detection.