Fault diagnosis of rolling element bearing is important to improve the performance and the reliability of mechanical systems. The extraction offeature paramet ers is essential to diagnose faults. The sample entropy was introduced into the field of fault diagnosis. Its performance and the choice of calculation paramete rs w ere discussed. Combined with wavelet packet decomposition and sample entropy, a feature extraction method for rolling element bearing faults was proposed. First ly, the bearing vibration signal was processed with wavelet packet decomposition . Then, the sub-band with largest normalized energy was reconstructed. Finally, the sample entropy of the reconstructed signal was calculated and used to evalua te the fault condition. The practical application proves that the method is effe ctive on fault diagnosis of rolling element bearing.