A new method of fault diagnosis of in-situ sampled vibration signals is introduced for the non-stationery signals of rotating machines by using the harmonic wavelet filtering. Based on the nice box-shaped characteristic of spectrum of harmonic wavelet, the components in prescribed frequency region and other frequency components are decomposed to independent frequency bands without any overlapping or leaking. Then the frequency components in the prescribed region can be extracted by data reconstruction. Simulation results and practical examples show that this method has good filtering effect. It can extract the valuable frequency components for elementary fault diagnosis. This method has the merits of simple algorithm and computer-time saving.