Aiming at the problem that the existing multi-manifold learning method does not consider the boundary information between the manifolds, which leads to the difficulty of data classification after dimensionality reduction, a new margin discriminant multi-manifold analysis (MDMA) method was proposed. In the MDMA, the intra-class similarity, inter-class difference, congeneric manifold structure and heterogeneous manifold structure of the data were considered at the same time, and in order to avoid the problem of small sample in the process of dimensionality reduction, the four points were summed up as the exponential trace quotient optimization structure when constructing the objective function. The experimental data set of two rotor systems was verified, and the results showed that compared with other typical dimensionality reduction methods, the method could extract the discriminant information contained in the data more effectively, and showed better classification performance in fault identification.