Recently, the complex network has been the rise of a new theory, which has rapidly permeated from natural science to the engineering science and so on. From the essential characteristics of the complex network community structure, fault samples are abstracted into the network nodes and the connection between samples is abstracted into edge, then we can establish the network model of fault data. We can make use of the concept of complex network node correlation to select community initial clustering center and Euclidean distance function to realize network initial division, design community distinguish criterion function, introduce the changes of modularity index to integrate the similar communities, and finally realize the accurate community division and fault diagnosis. We can apply the complex network community clustering proposed in the paper to these fault types which are difficult to distinguish to realize accurate classification. The proposed method applied in the examples of rolling bearing fault diagnosis verify that this method has a higher fault recognition rate.