Fault feature was always hidden by strong noise background in rolling element bearing fault signal. Based on morphological undecimated wavelet decomposition (MUWD), a novel approach was proposed to extract rolling element bearing fault feature. MUWD possess both the characteristic of morphological filter in morphology and multiresolution in wavelet transform. Signal length was maintained invariable and information loss could be avoided in MUWD. Multi-scale MUWD was developedbased on the characteristic of impulse feature extraction in difference morphological filter. The method was used to extract impulse feature in bearing fault signal. Experiment results showed that the presented method can achieve a better performance than traditional wavelet packet. MUWD algorithm includes addition, subtraction, maximum and minimum operations, and does not involve multiplication and division. It is suitable for on-line monitoring and fault diagnosis of bearing.