Research Of Fault Diagnosis Based On Support Vector Machine

Statistical Learning Theory (SLT) is a theory, which researches the machine study especially in small samples. Support vector machine (SVM) is a new machine learning method based on the statistical learning theory. It improves the algorithm generalization effectively and minimizes the empirical risk simultaneously by using Structural Risk Minimization and synthesizing the techniques including the statistical learning and neural networks.The superior performance of support vector machine to small samples attracts attention of investigators in fault diagnosis field. Fault diagnosis is a subject of small samples and also a problem of pattern classification in nature. The predominance of SVM applied to fault diagnosis is proper for small samples decision. The nature of the algorithm is acquiring connotative class information to great extent from small samples. From the point of generalization, SVM is more suitable for the practical engineering problem such as fault diagnosis.To begin with, the research content of fault diagnosis and its method, as well as the research status of fault diagnosis using SVM are introduced. What's more, this paper gives a specific introduction of the related knowledge of SLT, the principle of SVM and two main types of multi-category classification, that is,"one-against-rest"and"one-against-one". A method of fast feature selection is used in the fault data of a practical problem to reduce the number of its attributes, according to the mean and the square of each feature of data. Compared to the original solution, the simulation result shows that dimensionality reduction and computing complexity reduction were realized by using this method on the basis of achieving good performance of the multi-classifier.The research topic of this paper comes from the project of Intelligent Railroad Track Crack Detection System. The task is to judge the crack of railroad according to classifying the measured data using detection instrument. At first, we use neural network to analysis fault data and achieve the visualization of fault diagnosis system. After testing new sample data, we find that the generalization ability of neural network model is not very good. Therefore, SVM is introduced to this project and the"one-against-one"SVM is adopted. The simulation test achieves good classification results and shows that the superiority of SVM in dealing with the issue of small samples.Finally, a new algorithm called M-ary SVM (multi-category classification based on the binary classification version of SVM) is introduced to our project. After the simulation, this method can achieve good performance as well as"one-against-one"SVM, but the number of classifiers has been reduced.