A wind generating set fault prediction method based on D-S evidence fusion is disclosed. In the method, for two kinds of signals, two support vector machines after parameter optimization are constructed, the two support vector machines are taken as two evidences, and after D-S fusion, a final prediction fault type is given. The method has advantages that (1) in a traditional vibration method, only a vibration signal is analyzed, and a neural network, a decision tree and other machine learning algorithm models are constructed according to a vibration energy characteristic vector; but the vibration signal is only observed so that some fault states can be misclassified, for instance, a bearing damage and rotor eccentricity can cause a vibration signal abnormity, and at this time, a current signal can be used to distinguish two kinds of fault states; and (2) a prediction model established in the method can be stored, historical data does not need to be repeatedly extracted and trained, and under a real-time prediction environment of a wind field, a prediction result can be quickly provided.