Adaptive Wavelet Packet Feature Extraction Support Vector Machine Model And Spectral Analysis Applications

As a non-destructive analytical technique, Near-infrared (NIR) spectroscopy, a new high-efficiency analytical method, has been receiving considerable attention in many field due to its high measuring speed, zero pollution, very low running costs, simultaneous determination of components and requirement of less or even no sample preparations which it highly suitable for on-line process monitoring. However, because NIR spectra generally consist of broad, weak, non-specific and overlapping bands, the concealed relationship between the quantitative properties of samples and their NIR spectra, thus, are not too easy to be exhibited. Frequently, regression model is demanded in quantitative analysis of NIR spectra. The main aim of this thesis is to propose a new chemometric method for the analytical of Near-infrared spectra. The main achievements of this thesis are as follows:Combine the feature of Near-infrared Spectra, we develop a novel adaptively configured wavelet packet transform support vector machine as globally optimized by particle swarm optimization algorithms. Rather than wavelet transform only decomposing approximation coefficients each time, wavelet packet transform continuously decomposes detail coefficients as well. Actually, it is wavelet transform where original signal passes though more filters than wavelet transform. This treating makes wavelet packet transform more flexible in feature extraction, baseline and processing smooth signals like NIR spectra. Support vector machine is a learning machine on the basis of statistical learning theory principles. It is commended by its structural risk minimization principle and typically provides good generalization performance in many cases. To be really an attractive technique, SVM has been receiving increasing interest in numerous fields. The use of particle swarm optimization algorithms may allow synergetic optimization of the wavelet packet transform and support vector machine according to the performance of the total model, because the construction of decomposition tree in wavelet packet transform is in terms of a discrete optimization issue and parameter determination of support vector machine could be a continuous optimization issue. For the analysis of the near infrared spectrum of meat data, we found the quantitative analysis establish by this method is more robust compared with neural network, traditional support vector machine methods. The new algorithm is applied to quantitative analysis of the near-infrared spectral data of the corn’s moisture and protein. And compare with partial least squares, neural network, the wavelet packet transform of based on entropy minimization rule combined with support vector machines (SVM) algorithm. Although the particle swarm optimization based on the wavelet packet transform support vector machine (SVM) model also appeared certain a fitting phenomenon, compared with other model training set and forecast set the sample root mean square error value were decreased significantly, the accuracy of the prediction model and generalization ability has improved a lot. And provide a reliable method of near-infrared spectral technology combine chemometrics for the quality of corn.There is great significance of the whole animal husbandry for the analysis of the dry grass. The near infrared spectral technology in combination with chemometrics used for the quantitative analysis of the effective ingredients in hay, it is good for realize the rapid and low cost of hay quality analysis. In this paper, combining with near infrared spectroscopy and the adaptive wavelet packet transform feature extraction support vector machine based on the particle swarm optimization algorithm modeling, the main components of the hay:carbon, nitrogen and sulfur content was been quantitative prediction. The results show that this method can provide an effective way for improving the analysis level of the hay.