In view of the sliding bearing nonlinear and non-stationary vibration signals and its weak friction signals， a method of feature extraction and state recognition method for sliding bearing friction faults based on variational mode decomposition（VMD） was proposed. The VMD was used to adaptively decompose the sliding bearing vibration signals into the system shock signal， low frequency friction signal and high frequency friction signal. On the basis of VMD decomposition， the feature parameters of relative frequency spectrum energy moments were defined and extracted， which were used to describe the characteristics of sliding bearing vibration signals and their components. The crankshaft bearing friction fault signals from a S195-2 diesel engine were analyzed， and the average recognition accuracy of K-nearest neighbor classifier reached 93.3%. Results indicate that the relative frequency spectrum energy moments extracted on the basis of VMD are sensitive the working conditions of sliding bearing and can effectively identify its friction fault states.