Abstract:

In order to enhance the diagnosis efficiency with the accumulated network sample, and because the standard support vector machine doesnt support incremental learning directly, a boundary deflection overlay incremental support vector machine is proposed. According to the movement of separating hyperplane caused by the newly added training samples that violate KarushKuhnTucker conditions, the boundary deflection overlay algorithm is also designed to preextracts support vector reproducing region as the work set for incremental training, which solves the problem that nonsupport vectors transform to new support vectors. Analysis and simulation show that the method can not only reduce the work set effectively, but improve the training efficiency greatly without affecting the diagnosis accuracy.

Key words: network fault diagnosis, support vector machine, incremental learning, model update