Abstract: It is the first time to propose the application of wavelet decomposition to non-stationary time series forecasting. Non-stationary time series can be decomposed into several pseudo-stationary time series with wavelet decomposition. Each pseudo-stationary time series is forecasted with AR(n) modal to get the final result of forecasting. The method is used in forecasting wear trend of a beating pedestal in a compressor driving. system. Compared with Back-propagation network based method, the method obtains far more Precise results with shorter time, and can be applied to forecasting of machine running condition and analysis of machine fault trend effectively.