In light of the problem of inadequate use of prior knowledge and the difficulty of diagnosis in a high dimensional feature space, a semi-supervised spectrum kernel clustering method is proposed that combines pairwise constraint information and construction of kernel function through the constraint rules. First, using prior knowledge to construct the pairwise constraints, the kernel function can be designed by graph structure information and constraints information. Then, the projection matrix can be calculated. Finally, clustering can be operated in the projected space with a k-means algorithm. In the test set, every new sample can find k neighbor samples in the train set and the projected values of the k neighbor samples, and the local projection matrix can then be calculated. Thus, the projected values of each new sample can be calculated online. The experiments show that the proposed algorithm is more accurate in clustering than the comparative algorithms, and can meet the actual needs of mechanical fault diagnosis.