Abstract: The paper explores a way of extracting fault features from the vibration signals of multi-vibration sources, making use of the wavelet packet decomposition of the vibrant signal of a diesel engine crank-shaft bearing, thus reconstructs the time series of the wavelet packet decomposition coefficients in different frequency bands, and analyzes the time series by the AR (autoregressive) model spectrum, picks up the fault characteristic signals of the object analyzed. The result shows that the wavelet packet-AR model spectrum can separate the disturbance of multi-bestirring sources and pick up the fault characteristic signals of the diesel engine crank-shaft bearing effectively. The fault feature frequency band of the crank-shaft bearing is 0～0.25, which is more distinct at the engine's rotative speed of above 1 800 r/min. The optimal positions of the vibration sensor should be located on both sides of the crank-shaft bearing or bottom of the engine.