Study On Support Vector Machine Classification Algorithm And Its Application In Safty Evaluation Of Human Accident In Coal Mine

Support Vector machine (SVM), which bases upon the VC theory and the principle of structural risk minimization in statistical learning theory, has been a new approach in machine learning. It possesses advantages such as the solid mathematic theory foundation, well generalization, as well as seeking for the best balance between the model complexity and empirical risks. SVM not only can solve the problems such as over learning and dimension curse in traditional machine learning algorithms, but also has remarkable performance in dealing with the small-sample, high dimensional pattern recognition problems, which has already become the hot spot and difficult issue of research in machine learning field. Because of these advantages, SVM provides a lot of new conceptions for practical applications whose samples are quite difficult to obtain (e.g. the human accident in coal mine). Recently, with the continuously boosting demand of coal in our country, accidents have taken place frequently. According to the statistics, most of the accidents in coal mine are directly or indirectly caused by human accidents which have become the major factors of the security of the system. However, due to the particularities of coal mine industry, miscellaneous forming factors in human accidents, the hardness in feature extraction, the complexities of working mechanism, the difficulties of statistical analysis, as well as the rarely available samples, there dwells certain limitation in the assessment of human accidents by only using traditional methodology of safety system engineering. Nevertheless, SVM provides the effective way to resolve the small-sample, high-dimension and non-linear human accidents evaluation.By taking statistical learning theory as the main line, this paper analyses the shortcomings in SVM classifier, and investigates the fusion methods between SVM classifier and other theories, such as manifold learning, kernel density estimation, similarity margin of inner product, kernel alignment, etc. Then, based on the improvements, this paper combines the SVM classifier with the theory of coal mine safety assessment, and constructs the evaluation system of coal mine human accidents. This thesis mainly focuses on the following aspects:1. The research on the improvement of SVM classification algorithmAiming at settling the shortcomings in SVM such as it is sensitive to the distribution of samples near the separating margin, this paper utilizes manifold learning algorithm to investigate the inner law of sample distribution, and combines with the spread directions of manifold to analyze the changing trends. We also study the evaluation approach of classification under the nonlinear mapping from observation space to intrinsic embedding space, to construct the objective function to train the classifier.To resolve the hyperplane corresponding to the largest margin, SVM has to calculate the sophisticated quadratic programming problem, which is time-consuming. This paper, which associates kernel density estimation with SVM, proposes a novel direct SVM classification algorithm. And with advantages of simple model and less calculation, the new classifier could maintain well generalization performance as well as better classification accuracy.Besides, on account of the limitations of incremental learning in classical SVM, an incremental PSVM learning algorithm based upon the cloud model is advanced in this paper. The fast learning ability of PSVM is exploited to yield the initial classification hyperplane, and then, all training datasets are reduced through using k-NN method and the plane. After that, we apply the cloud model to directly carry out discriminate analyses on the reduced dataset. The simple algorithm could well embody the distribution of incremental samples and be solved without iteration. Experimental results manifest that this algorithm could not only keep well classification accuracy and generalization ability, but also improve the training speed.2. The study on the evaluation index system of coal mine human accidents.After discussing the theory of safety evaluation in coal mine in depth, this paper has made elaborate analyses of leading factors in coal mine human accidents. Then, according to characteristics of human accidents, the coal mine human accidents'evaluative indices and relevant feature extraction method have been come up with in this paper, which convert the feature selection of intricate multi-factors to the parallel combinatorial optimization and the global optimization. Through the effective feature selection, structural complexity of the system could be decreased, and the interference imposed by redundancy factors could be eliminated. After that, this paper analyzes application of the improved SVM models in human accident assessment.3. Studies on the model selection algorithm of SVM and its application in influencing factor analyses of coal mine human subsystem reliability.The selections of optimal kernel function and its parameters are of great importance in SVM. From the point of view of margin, this paper further investigates the relationships among generalization performance, VC dimension and margin in SVM, proposing a novel similarity margin of inner product. Then, the richness or flexibility of the corresponding function class, along with the estimative upper bound of relevant dimensionality, is analyzed in detail. Meanwhile, by investigating the kernel density estimation theory, this paper constructs the evaluation index of kernel parameter model in high dimensional feature space based upon kernel density estimation, and then, puts forward a model selection approach of kernel parameters for SVM based on the kernel density estimation.On these bases, this paper analyses the reliability and influencing factors of human subsystem of fully-mechanized mining faces in detail. Meanwhile, we employ the kernel density estimation clustering to construct a multi-index reliability grade division method of human subsystem, and then, the mapping between influencing factors and the reliability grade is established by SVM. By using our model selection approach, this division method has better nonlinear capability.4. The research on multi-class SVM and safety assessment model of organization and management factors in coal mine.A new multi-class approach to support vector machine classification, by which the parameters can be effectively reduced and the complex quadratic programming can be avoided, is proposed, settling the limitation of traditional SVM in the multi-class classification problem. Then, this multi-class algorithm model is applied to analyze organization and management factors in coal mine. From perspectives of human management, organization or institution management, enterprise environment management, field and technological management, this paper analyses the main influencing factors of organization and administration of coal mine in detail. After that, based on the kernel alignment and SVM, a method for predicting the safety evaluation grade of organization and management is presented. At the same time, relationships between the safety evaluation grade of organization and management and influencing factors are also analyzed.5. The analysis of the prototype system of human accident safety assessment in coal mine.Finally, this paper analyzes the design and implementation of the prototype system of human accident safety assessment of coal mine. Through combining the theory with practice, we discuss the key issues in system design from the software engineering side. This system, which could improve the safety management in coalmines and provide the consult and reference for the development of other system platforms in the relevant fields, is practical.