In order to diagnose fault effectively by using sensitive features contained in the feature set, KFDA was improved in this paper and a fault diagnosis method based on improved KFDA individual feature selection was proposed. Firstly, the mixed feature of the fault vibration signal was extracted from different angels, and the original high-dimensional and multi-domain feature set was constructed. Then, an improved kernel Fisher feature selection method was proposed and used to select individual sensitive feature subset for each pair of class. Finally, a one-against-one approach was applied to train several relevance vector machine（RVM） binary classifiers, and sensitive feature was input into the multi-class fault diagnosis model for recognizing the fault types. The experimental results of gear indicate that the proposed method is of high diagnostic accuracy.