Abstract:  Blind source separation was used to analyze non-linear vibration of RC beams. The algorithm called second-order blind identification was applied to the structural vibration signals de-noised by wavelet soft-thresholding to acquire the first order vibrational component, which was then processed by Hilbert transformation to get the frequency-amplitude curve. Based on the verification by finite element simulations, the recorded vibration signals of a RC simple beam and a prestressed simple box-beam were analyzed by the above procedures. The results show that blind source separation has good ability of anti-noise and is better than short time Fourier transformation regarding ease of implementation. In addition, the damage identification by non-linear vibration analysis is reliable than that by frequency change obtained by traditional modal analysis based on linear vibration.