| Pattern classification method is one of the key techniques which are widely used in many engineering fields. With the development of science and technology, more and more chemical pattern messages are obtained, so the dimensionality of a chemical pattern is higher. Because the measurement of some data of a chemical pattern is costly, usually the size of pattern matrix is not big enough and the collected specimens do not obey a known distribution, and moreover there exists the multiple-collinearity among the elements of patterns. Therefore the chemical pattern classification is more difficult than before.In this thesis, the advantages and disadvantages of multivariate discriminant analysis and feed-forward neural networks as classifiers are discussed, and then the strategy integrating the classification methods of two different kinds is proposed. According to this strategy, the new integrated classifiers are established and then are applied to two chemical pattern classifications after analyzing their classifying ability.Discriminant analysis is one of the classical classification methods based on multivariate statistical analysis. The process of calculation is canonical and each step has its own probabilistic sense. But unfortunately the collected specimens must obey a certain distribution and the sample size should be big enough. In complicated chemical pattern classification problem, the result isn't always so good when only discriminant analysis method is applied.Feed-forward neural network (multi-layer feed-forward neural network and radial basis function neural network) has the excellent intelligence of calculation and is able to efficiently approximate a function of any form, so it is suitable for the nonlinear classification. Exit it is of great difficulty to determine the structure of a feed-forward neural network and usually a bad structure leads to over-fitting and under-fitting. Moreover, if an inappropriate training algorithm is selected, the process of determining the parameters of a feed-forward neural network is time-consuming and possibly stops at a certain local minimum. By improving the ability of approximating and precision of training, the classification correction of training set reaches a high value, but that of testing set increases in a small degree or sometimes decreases.Sequential discriminant analysis (SDA) is a classical method to select variables.Ill?(ABSTRACT)When SDA is applied in the sample of nature spearmint essence, some unimportant variables are removed and important variable is remained. Principal component analysis (PCA) and correlative component analysis (CCA) are commonly used techniques to extract components. By comparison, the effect of classification correlative components identified by CCA is better than that of principal components in chemical pattern classification.The main focus of this thesis is on combing the feed-forward neural networks with traditional statistical methods as well as putting forward the integrated pattern classification strategy. Feed-forward has great ability of approximating, and moreover the specimens don't need to follow a certain distributing. Using the self-learning of feed-forward neural networks, the original patterns are mapped into a new pattern space where the distributing of specimens is helpful to classification. Because there exists correlation between the elements of a pattern, so CCA is applied to extract useful components, which are vertical to each other, from the transformed matrix. At last, a discriminant model is established.The main process can be described as follows:(1) Derived from the multi-layer feed-forward neural network, the WS.T transform, which is able to project the original patterns vector into a new pattern space, is established.(2) Based on the virtue of the radial basis function neural network, the RBF.T transform is designed to realize the pattern conversion.(3) By applying the traditional statistical component extracting method, important components of classification...
|
PDF Full Text Request