The simulated vibration signals of a simply supported beam are decompo sed into a series of intrinsic mode functions (IMF’s) by using the empirical mo de decomposition (EMD) method, and then the energy percentage of the first four IMF’s are evaluated and used as the input parameters of an artificial neural ne twork model which is trained on a dataset with 136 pieces of data extracted from 17 types of beam faults. The neural network model is tested on a separate valid ation dataset, which reports all the faults correctly. As long as the fault data base contains enough fault cases, the joint method can detect fault statuses, fa ult types and locations.