The faults of the friction hoist spindle usually show characteristics of coupling, weak feature and less availability of samples. In light of these problems, a fault diagnosis method based on complex network clustering was proposed. Starting from the essence of the community structure, which was displayed by fault data, a weighted and undirected complex network model was constructed by nodes that abstracted from each sample and weighting edges that were represented by the similarity of samples. The generalized Ward distance, which was obtained by extending the concept of distance from Euclidean space to a similarity measurement, was proposed as a distinguishing criterion. Then, a clustering algorithm of a network model was developed by a hierarchical and agglomerative progress, namely, the pattern recognition of fault samples was accomplished. By analyzing the fault samples acquired from overload, elements failure of rolling bearing, and worn gear of the reducer, the experimental results indicate that the proposed method can effectively cluster fault samples of a known type and provide data support for collecting unusual samples, which can be used to discover and diagnose the unknown fault pattern, for the number of types was not preset. The proposed method is more accurate and productive than multi-class support machines and Fast Newman algorithm.