The existing complex network methods are directly from the time domain when it is applied in fault diagnosis, which causes frequency domain information of the signal missing, and the extracted topology features of network are too macroscopic, which is not sensitive to network within the local change. Meanwhile, local features usually have more abundant information and represent the network model more accurately than macro features. As a consequence, a new method of local feature extraction based on frequency complex network decomposition is proposed. The method obtains the changing rule of the signal in the frequency domain with the aid of the structural characteristics of complex networks and uses the microscopic features that is sensitive to network within the local change to represent the whole network. It is flexible to apply without limit by the mechanism. Classification experiments on different faults of rolling bears are conducted to compare the proposed feature, existing complex network topology features and statistical parameters in time domain. The experimental result indicates that the proposed feature has well separability and high recognition efficiency, which could satisfy the need of rolling bearing fault diagnosis and also has a reference value for the non-stationary signal processing of other parts in machine.