A fault diagnosis method based on KNN evidence fusion is presented to keep a balance between modeling accuracy of fault feature data and instantaneity of diagnosis decision making. For each fault feature(symptom), its historical sample data are used to model fault template patterns(FTPs) with the form of random-fuzzy variable(RFV), the KNN algorithm is used to find out K historical samples nearest to a testing sample and the RFV-type fault testing patterns(TPs) of the K samples are presented to describe the testing sample. The matching degree between FTP and TP can be calculated to generate the K pieces of diagnosis evidence, and then all evidence coming from the different fault features can be fused and diagnosis decision can be made based on the fused result. In this method, the fine modeling can be realized by using the RFV, and meanwhile, the diagnosis information of the single testing sample can be enriched by adding the K historical samples, and the instantaneity of diagnosis can be improved. Finally, in diagnosis experiments on a rotor test bed, the effectiveness of the proposed method is verified.