Aiming at the problem that the dimension of faulty data set with strong nonlinearity in fault diagnosis is too high and the sample of faulty label is insufficient， a kernel semi-supervised local Fisher discriminant analysis （KSELF） is proposed. The method first maps the original fault data set to a high-dimensional feature space by the kernel method， and then derives the projection transformation matrix based on the semi-supervised local Fisher discriminant analysis in the high-dimensional space. The proposed KSELF can effectively capture the nonlinear information of the data， and can fully utilize the fault information in a small number of label samples and a large number of unsigned fault samples， thereby avoiding over-learning problems. The proposed method is validated by the fault characteristic data set of a double-span rotor test bench. The results show thatthe KSELF method has stable dimensionality reduction ability， and can obtain better dimensionality reduction effect and higher classification accuracy， compared with other methods in the experiment.