To improve the efficiency of fault diagnosis, a novel granular computing algorithm is developed to reduce computational cost. It is realized by extracting and partitioning on the complete graphs, and in the process of graph generation, the graph partition based on dimensionless similarity (GPDS) method is proposed to overcome the influence of attributes with different dimensions. The algorithm is named graph partition based on dimensionless similarity. Moreover, similarity threshold determination method based on frequency distribution histogram is proposed to reduce the dependency on the experiences of experts. Meanwhile, a weighted relative error is proposed to measure quantitatively the distribution change of original data after being compressed. Finally, different characteristic data are applied to verify the theories. The experimental results indicate that the compressed training samples can maintain the classification accuracy. Furthermore, the elapsed time can be obviously reduced. Therefore, the GPDS method can be used in fault diagnosis properly