In the light of marked fault samples and low recognition rate of the traditional semi-supervised diagnosis method，a novel fault diagnosis method, which is based on semi-supervised max-margin dictionary learning (SSMMDL), is put forward. In the proposed method, an unmarked sample reconstruction error term is added to the max-margin dictionary learning model. By minimizing the four terms, including unmarked samples sparse reconstruction error term, marked samples sparse reconstruction error term, the regular items of the loss function in support vector machine (SVM) and the regular items of classification interval regularization, the synchronous learning of the dictionary and the SVM is realized. and the dictionary with the discriminant ability is built. On this basis, the sparse representation of test samples is obtained. At last, we identify the faults by the classifier based on sparse representation. The different faults of rotor are recognized. The experimental results show that the proposed method is more accurate than the comparative algorithms in terms of accuracy rate, and succeeds in meeting the needs of online monitoring of mechanical fault diagnosis.