Virtual Sample And Its Application In The Coal Maceral Classification

Classifying and recognizing coal macerals, building the relationship between macerals parameters and coal properties accurately have important significance for efficient and clean utilization of coal. However, rich samples of some macerals are not so easy to obtained that makes classification difficult. Research showed that the virtual sample technigue is an effective means to extend the training sample set and improve the generalization ability of learning machine. In view of this, based on the analysis of coal microscopic image information characteristics, Gaussian distribution based virtual sample algorithm is improved and employed to expand sample set of coal macerals, and to improve the generalization ability of classifier. The main contributions of this dissertation are as follows:(1) After reading related literature, research status of coal maceral classification and virtual sample generation methods are summarized.(2) With an example of vitrinite, features are extracted and selected, including 5 gray- level histogram based features and 16 texture related features based on gray level co-occurrence matrix and run- length matrix. After analyzing the distribution characteristic of each maceral, feature amounts which not only meets the Gaussian distribution but also can effectively distinguish each maceral are selected, and a feature parameter set is built.(3) By using the maximum likelihood estimation, the original Gauss virtual sample method is improved. With maximum likelihood estimation method, parameters as mean and variance are estimated, a probability density function(pdf) of Gauss distribution is constructed. Then, a certain number of Gaussian samples around the initial samples are randomly generated based on the constructed Gaussian distribution, which are mixed to the initial samples to generate a new training set.(4) In order to test the influence of initial training date number and virtual sample number on classification, 5, 10, 15 and 20 initial samples, named set ?, were selected from each maceral. And 5, 10, 15, 20 and 25 virtual samples are added to the initial training set ?, respectively, to format a new training sample set, named set ?. Then, a radial-basis function support vector machine classifier, respectively trained support vector machine on training dataset ? and set ?s, and total classification accuracy are obtained with another testing sample set.(5) The above method is applied to the other two macerals: inertinite and exinite, the influence of virtual samples on the classification result and the adaptability of the algorithm are analyzed.The main characteristic and innovations of the dissertation lie in: To the difficult of poor samples of some macerals, a virtual samples method is introduced to expand the training sample set. According the statistical characteristics of maceral features, the maximum likelihood estimation is used to obtain the parameter of Gauss ian distribution, to make some improve of original Gaussian virtual samples method. The total classification accuracy and the generalization ability of the classifier are improved.