In this paper, a cluster-based feature extraction from the coefficients of discrete wavelet transform is proposed for machine fault diagnosis. The proposed approach first divides the matrix of wavelet coefficients into clusters that are centered around the discriminative coefficient positions identified by an unsupervised procedure based on the entropy value of coefficients from a set of representative signals. The features that contain the informative attributes of the signals are computed from the energy content of so obtained clusters. Then machine faults are diagnosed based on these feature vectors using a neural network. The experimental results from the application on bearing fault diagnosis have shown that the proposed approach is able to effectively extract important intrinsic information content of the test signals, and increase the overall fault diagnostic accuracy as compared to conventional methods.