A new way of extracting fault features from the crankshaft bearing fault of diesel engine by using the empirical mode decomposition (EMD) method and auto regressive (AR) model spectrum analysis technology is proposed. By using EMD the unstable vibration signal of the engine at the stage of speedup is decomp osed into a series of stable intrinsic mode functions (IMF). The first 5 orders of IMFs taking up the main parts of the origin signal are analyzed by the AR mode l spectrum, the frequency band power of the AR model spectrum of IMFs is calcula ted and the power which reflects the fault of crankshaft bearings is extracted as the feature of the crankshaft bearing fault. 6 vibrating sensors at different locations and data acquisition unit running at 4 different speeds of the engine are set, and the acquired vibration signal is analyzed by the proposed method. When the sensor is located on the right joint between cylinder and oil panats peeds of engine greater than 1300r／min, the extracted features are highlighted. The result shows that the extracted fault features can exactly reflect the fault of crankshaft bearing.