A new feature extraction method for localized faults was proposed in order to distinguish fault features in a single channel vibration signal. This paper combines the concepts of time-frequency distribution with non-negative matrix factorization to propose a novel time-frequency matrix factorization method to extract representations of localized faults. The short-time Fourier transform was adopted to describe the localized faults of the vibration signal, and the subspace dimension was decided with a consistency index. Then, based on the decomposition of the time-frequency spectrum with orthogonal constraints, the low dimensional embedding was effectively separated. Finally, the theory of phase recovery was adopted to reconstruct waveforms of the localized fault features of interest. The simulation and rolling element bearing results showed that the proposed method was effective at extracting fault features of a single channel vibration signal and improved recognition accuracy in a mechanical system.