In order to improve the fault diagnosis accuracy of the complex system of diesel engine, based on the acquisition of the characteristic information of the sound emission signal of the cylinder head, the vibration signal of the cylinder head and the acoustic emission signal of the cylinder block, three BP neural networks are used for local diagnosis and obtaining the evidence bodies and D-S evidence combination rules are applied to each evidence body for decision fusion. In view of the limitations of D-S evidence theory in dealing with conflicting evidence, a fault diagnosis method based on multi-source weighted fusion is proposed. The correct rate of recognizing all the propositions and the distance between evidence bodies are used to correct each body of evidence. The fault simulation experiment shows that the proposed method can improve the accuracy of diesel engine fault diagnosis and fully verify the feasibility of its application in diesel engine fault diagnosis.