An indirect approach for distinguishing chaos, surrogate data method, is proposed and applied to analyze non-stationary fault signals in this paper. Firstly, the theory and algorithm of surrogate data method is investigated. And then the correctness and availability in the analysis of the method is tested by three kinds of typical known time series. Finally, we utilize the approach to analyze two groups of non-stationary fault signals from rotor system. The results show that this method is an effective one to identify chaotic characteristic of non-stationary fault signals quantificationally. It can also provide a new way for nonlinear feature etraction and recognition of rotor system typical faults.