In order to improve the accuracy of vibration fault diagnosis for hydroelectric generating unit, a method is proposed based on quantum particle swarm optimized BP neural network (QPSO-BP) which is fused by modified D-S evidence theory. According to the common vibration fault types, three QPSO-BP networks with inertia weight being adjusted randomly are used as its primary diagnosis models, then the independent bodies of evidence are applied to the synthesis of D-S theory, and the basic probability assignment is realized objectively. In view of the defects that standard D-S can not synthesize high conflict evidence, the weight matrix is calculated to improve it. Example analysis shows that the proposed method can effectively identify vibration fault of the unit, and has a high diagnostic accuracy compared with three primary diagnostic model and standard D-S theory.