A novel feature extraction technique based on the morphological liftin g wavelet (MLW) decomposition was presented for roller bearing fault diagnosis i n this paper. Application results of simulated impulsive signal and the real fau lt bearing vibration signal proved the effectiveness of the MLW. Compared with t he traditional linear wavelet technique, the MLW, which demonstrated to be more available to conserve the impulsive signals and retaining noises, can identify t he bearing fault which the linear wavelet failed to detect. Another attractive c haracteristic is that only simple addition, subtraction and comparison operators are used in the MWL, so the cost of computation is extremely low. Thus the MWL can be applied to the online condition monitoring and fault diagnosis of roller bearings.