The Application Of Wavelet Transform In Specturm Data Processing

Denoising is an important part in the spectrum data processing. Various conventional methods had been developed in the field, but they are not fully satisfactory regarding precision and efficiency. Wavelet transform can analyze signals at time domain and frequency domain simultaneously, so it can denoise effectively. Another important problem is the threshold selection. The threshold selection has immediate relation to the result of denoising. Partial wavelet cofficients can not be set zero when the threshold is undersize so that parts of noises are retained. Some useful signals will be taked off when the threshold is overestimate so that parts of useful signals are lost. So how to select threshold effectively to avoid lose useful signal is a problem worth to research.The works developed in this field by former researchers were summarized. The excellent properties of wavelet transform were then explored in denoising Near-Infrared Spectroscopy data associated with software tools such as MATLAB and VC++6.0, and programmed an application of denoising by C++ program language. The contents of this thesis were organized as follows:Chapter one is exordium. In exordium the research purpose and means is introduced. Then the research of this study was introduced. At last are the tools of the study.Chapter two reviews the basic theory about wavelet transform. In comparing with the Fourier transform, the properties of localization and adaptive time-frequency window of wavelet transform were interpreted in detail.Due to the redundancy of continue wavelet transform, the preference of using discrete wavelet transform was also explained. The principle of mutile-resolution analysis was dicussed, and the decomposition of signal frequency in space-scale was described. Mutile-resolution analysis was used to explain the principle of discrete wavelet transform.Chapter three introdued the theory of denoising by wavelet transform. The various noises and noise removing methods including the method based on wavelet transform was extensively discussed. Attention was paid on the noise pulse with its width close to that of a signal peaks. It is important but diffcult to identify this kind of noise pulse from signal peaks in spectrum data automatically. The spike of a normal spectrum peakis usually smoother than that of a noise pluse caused by imperfection of experimental running, so the singularity analysis through wavelet transform was proposed to identify whether the spikes correspondingto the spectrum peaks or noise pluses. And the several threshold denoising methods were introduced, and studied the result in simulate signal.Chapter four introduced wavelet transform denoising method used in application of actual spectrum data processing which included milk components measurement using Near-Infrared spectroscopy and blood glucose concentration by short-wave near infrared spectra. In the first experiment, Wavelet packet analysis method was used in data denoising processing of 38 transmittance spectra (collected by a Nicolet Nexus 870 FT-IR-Spectrometer) of pasteurized milk. After wave filtering, a changeable size moving window partial least-square regression method (CSMWPLS) was used for fat, protein and lactose model building. The accuracy of the predictive model was estimated by correlation coefficient (R) and root-mean-square prediction error (RMSEP). The RMSEP of fat, protein and lactose model is 0.0327, 0.0239 and 0.0631mg/L, respectively. By comparing the result of wavelet packet analysis with normal method, we can get the result shows that the wavelet packet analysis can denoising the signal effectively so related the targets better and thus build a good milk component predictive model. The second experiment is about wavelet analysis method was used in data denoising processing of 14 volunteers'18 serum samples' short-wave near-infrared absorption spectra from 700nm to 1060nm. Using the blood glucose value measured by glucose meter as standard value, iPLS(interval Partial Least Square) method was used to make the glucose prediction model. The accuracy of the predictive model was estimated by correlation coefficient (R) and root-mean-square prediction error (RMSEP). The result is R=0.9654, RMSEP=0.2435mmol/L. By comparing the result of wavelet analysis with normal Fourier transformation method, we can get the result shows that the wavelet analysis can denoising the signal more effectively and thus build a better predictive model.Chapter five is about the software of wavelet transform in denoising. Using C++ program language and building in VC++6.0 environment, we programmed a one-dimension wavelet transform applicational software. The threshold selection used coefficient weighting method. The source codes of wavelet transform have a real important applied value in software of spectram analysis.Chapter six is the summation of the work. The summation is carried out on the research, and the problems existing at wavelet denoising method research and practical application are discussed.