Abstract: To improve the early friction fault diagnosis capability, the singular value decomposition method was introduced. By utilizing the characteristic of strong periodicity of rotating machinery vibration signals, a Hankel matrix was constructed from the signals collected in a full cycle. The singular values of the matrix were obtained through the singular value decomposition method, and the singular values were divided into a large-value region, a medium-value region and a small-value region. Results show that the main harmonic periodic components in the signal can be reconstructed according to the singular values in the large-value region, the noise components in the signal can be reconstructed according to the singular values in the small-value region, and the signal reconstructed from the singular values in the medium-value region can monitor the periodic impact phenomenon, which is a typical feature of early friction faults. Combined with the analysis of a friction fault case, the impact feature of twice per cycle is effectively extracted, and the instantaneous energy impact phenomenon is further extracted, which verifies the feasibility of this method for early friction fault diagnosis.