To address the issue of neighbor-structure distortion in cross-modal fault data, a bearing fault diagnosis method based on CNCSL (cross-modal neighbor complementary space learning) was proposed. By this method, neighbor-structure complementarity among different modalities was achieved, and the accuracy of bearing fault diagnosis was effectively improved. Aiming at the complementary dilemma caused by inconsistent dimensions of data among different modalities, the implementation methods of space learning were considered in a reverse manner. A space-transformation strategy between the original data space and the cross-modal neighbor complementary space was designed, and cross-modal neighbor complementary cluster centers for all modalities were constructed. Thus a more authentic neighbor structure of the original data was obtained in the cluster centers, and the cluster centers was not affected by the inconsistent dimensions of the cross-modal data. Then, on the basis of the cluster centers, cross-modal neighbor complementary scatters were constructed to embed a more authentic neighbor structure into the cross-modal neighbor complementary space, which can enhance the neighbor-aggregation property of the space. Finally, by minimizing the cross-modal neighbor complementary scatters while maximizing the pairwise correlations of all modalities, the objective optimization function of the CNCSL model was constructed. The analytical solutions of this model were theoretically derived, and then cross-modal neighbor complementary features were directly obtained through space projection. These features not only preserve a more authentic neighbor structure as much as possible, but also exhibit good discriminability. The effectiveness of the CNCSL method is demonstrated by experimental results on the Paderborn bearing dataset and laboratory experimental platforms.