The manifold learning algorithm is a method of nonlinear data dimensionality reduction and can get available low dimensional geometry,it is also better reflects the nature of things. In order to improve the classification ability of gearbox vibration signals,a new feature extraction method from fault signals of gearbox vibration based on the manifold learning method（ MLM） is proposed. The results show that the fault feature information of gearbox vibration can be MLM effectively extracted and the different types of fault feature information can be separated. The diagnostic accuracy of features extracted by MLM is higher than that of the time domain statistical method.