Abstract: Fault tree model based diagnosis divides the bottom events of the fault tree into three parts: certain fault sources(CFS)、normal event sources(NES) and possible fault sources(PFS), but it is not clear how to determine the states of the elements in PFS (normal or abnormal). With lots of training examples, learning diagnosis method based on neural network is applied to determine the states of the elements in PFS, and its effectiveness is demonstrated by diagnosing a principle fault simulation testbed of any satellite power system.