Aiming at the problem of feature extraction and classification of weak fault information of the roller bearing，this paper proposes an Interactive Deviation Support Matrix Machine（IDSMM）. In the process of IDSMM modeling，the deviation parameter is added to each objective function. Geometrically，the hyperplane corresponding to each class of samples is deviation from the original position to construct a pair of interactive hyperplanes with deviation，and each hyperplane is further away from the heterogeneous samples，which improves the generalization ability of the model. In addition，IDSMM introduces a multi-rank left and right projection matrix to construct the objective function，which fully excavates the structural information between rows and columns in the matrix，and makes the model have better data fitting ability. Two data sets of roller bearings are used for experimental verification，the experimental results show that IDSMM has excellent classification performance in the fault diagnosis of the roller bearing.