Aiming at the problems of roller bearing fault diagnosis

gray theory and auto-regressive combination forecasting model is put forward

and the combination model has been build. The methodology developed decomposes the signal in intrinsic oscillation modes first

to translate the non-stationary signals into stationary signals. Then the autoregressive （AR） model of the selected IMF is established. The rough trend of the wear particle content change can be reflected through gray theory

and the detail of the change can be reflected through auto-regressive model. By testing and comparing a set of graphic data

the result shows that the combination model has a better forecasting result.