Abstract: To solve the shortages of Relevance Vector Machine (RVM) in fault prediction for largescale samples, such as the fall of prediction accuracy, low computation efficiency and so on, a method of Fuzzy Granular RVM (FGRVM) used for interval prediction of gradual fault parameters was proposed. Firstly, the original time series of parameters were granulated by fuzzy information granulation (FIG), and then the redundant data in the training set was reduced by a reduction approach based on the exponent similarity. Secondly, the Simple Particle Swarm Optimization (tsPSO) algorithm was improved by adaptive extremum disturbed and adaptive mutation, and the kernel width of RVM was selected automatically by the improved tsPSO (itsPSO) algorithm with the optimization target to minimize the error of 4fold cross validation. Finally, the changeable intervals of gradual fault parameters were predicted by the welltrained RVM. The prediction performance of the method was proved by simulations on three classic largescale, nonlinear time series with noisy data and Delta Exhaust Gas Temperature (DEGT) of an aeroengine in Air China, and the experiment results show that, compared with RVM and Fuzzy Granular Support Vector Machine (FGSVM), the prediction accuracy and computation efficiency are improved greatly by the method proposed in the paper.


Key words: gradual fault, interval prediction, FIG, RVM, reduction, PSO