Article Preview
Top1. Introduction
With various kinds of intelligent classification methods studied, support vector machine has recently been widely applied in the fault diagnosis of many engineering fields due to the unique advantages when faced with small sample data of faults and highly nonlinear relations between monitoring data and the actual system status (Saravanan, N., 2010; Guo, C., 2010; Wu, Q., 2010; Cui, H., 2009; Salahshoor, K., 2010). However, the disadvantages of SVM are also obvious including that the number of kernel function will increase significantly with increasing samples, the configuration of kernel parameters and regularization coefficients is complex and the kernel function must meets the Mercer conditions.
Relevance vector machine (RVM) is a sparse probability model based on Bayesian learning theory which has a similar function form with SVM and better generalization performance (Tipping, M. E., 2000; Tipping, M. E., 2001). Furthermore, it does not have to satisfy the Mercer conditions and has higher model sparsity. And the model is capable of outputting probability distribution with prediction results, and widely applied to engineering fields such as sewage detection (Zeng, T., 2013), hyperspectral image classification (Demir, B., 2007; Zhao, C., 2012; Yang, G., 2010), fault diagnosis and prognostics (Huang, K., 2010; Moghanjooghi, H. A., 2012) and so on nowadays.