Aiming at the non-stationary and nonlinear characteristics of fan vibration signals, a method based on time domain signal analysis combined with the improved k-means clustering algorithm is proposed. In order to estimate the fault types, peak to peak values of several typical fan fault signals, Hurst exponent and approximate entropy have been extracted and put into the improved k-means clustering classifier as feature vectors. The experiments on the centrifugal fan show that: the selected three kinds of time-domain characteristics can reflect the difference between faults and the effect． Besides, the improved k-means clustering algorithm whose average recognition rate may come up to 88.67％ has a better classification performance compared with the originalk means, and runs much more stably.