The invention discloses a bearing fault prediction method, and the method comprises the following steps: S1, extracting a bearing vibration signal feature vector from a bearing vibration signal; S2, deleting the signal interference in the bearing vibration signal feature vector through grey relational analysis, and obtaining a screened bearing vibration signal feature vector; S3, building a prediction model according to the screened bearing vibration signal feature vector, and outputting a fault prediction feature vector; S4, selecting a training sample from the fault prediction feature vector, and constructing a fault discrimination binary tree according to the training sample. Through the improvement of the prediction model, the method improves the adaptability of the prediction model, prevents the old bearing information from causing the old parameters of the prediction model, improves the prediction precision of the prediction model, and improves the recognition precision of a bearing fault.