According to the characteristic that a large number of redundant information existed in the fault diagnosis of ball screw, mean impact value (MIV) method was introduced to select features of trouble signal. Using this method, redundant features could be removed and the features that had greater impact on the diagnostic result were taken as the input of support vector machine (SVM). Then SVM was used to train input parameters and complete fault pattern recognition. By experiments verification, the ball screw fault diagnosis model has a better diagnostic result which has shorter time and high accuracy than before.