Abstract: Aiming at the shortages of low learning efficiency and overfitting of relevance vector machine (RVM) in the prediction of largescale fault data with RVM, the theories of granular computing (GrC) and RVM were combined systemly, then a fault prediction method of granular relevance vector machine (GRVM) was proposed. Firstly, taking fuzzy C-means clustering (FCM) as the granulation method, then by FCM, the original data set was granulated into several granules and replaced by the clustering centers. Secondly, the clustering centers set was taken as the training set to train the model of RVM. Finally, we used the trained model to predict the unseen data. The simulation results indicate that, the suitable selection of clustering number and kernel parameter is the key to ensure the performance of GRVM, and GRVM can improve the learning efficiency while it avoids the overfitting and keeps the high prediction accuracy.

Key words: GRVM, GrC, RVM, FCM, clustering number, kernel parameter