Research On Network Fault Detection Based On SVM

With the popularity of the Internet and the continuous expansion of its scale, network fault detection method by data driven is becoming the focus of network fault detection. This method uses data mining, artificial intelligence and other technologies through the analysis and processing of a large number of log files and historical data, to realize discovering network fault intelligently and positioning rapidly. In all data driven methods, support vector machine(SVM) is not only for solving a few samples, nonlinear and high dimensional pattern recognition well, for the larger data sets, the effect is also very good. Therefore, there is nice theoretical foundation and good application prospects that the support vector machine is used for network fault detection.The concept of similarity is proposed based on the characteristics of support vectors and nonsupport vectors in support vector machines; also the computational method of similarity is given. According to the similarity, an algorithm named Quickly SVM Based on the Similarity(SQ-SVM) is proposed. The algorithm simplifies training set through the value of similarity, and it is verified on the public datasets named cod-ma. Then, a new, simple, effective increment SVM method named Increment SVM Based on the Similarity(BS-ISVM) is brought up by combining generalized KKT conditions and similarity theory, and the increment method is proofed on the simulated datasets. At last, a network fault detection system is designed and implementation on the public datasets named KDD CUP99. After the processing of quantize, formatting, standardization and reducing dimensions for original data, the results of supervised learning and baseline learning(unsupervised learning) are compared and analyzed through SQ-SVM and One Class SVM.Experiments show that Quickly SVM Based on the Similarity and Increment SVM Based on Similarity can improve the training speed in case there is high enough accuracy. In network fault detection, it is necessary to preprocess the original data. Through comparing the normal support vector machine and one class support vector machine shows that the effect of supervised learning is better than that of unsupervised learning.