To cope with the shortages of low identification precision of mode shape of traditional weighted spectral density (WSD) method when select the exponential factor of weighting function inappropriately, a new mode shape identification method based on the WSD was established under frequency domain to further improve the identification precision of WSD and expand the application range. On the basis of the similarity between blind source separation method and mode superposition method, the structural vibration signal was decoupled and separated based on WSD method. An three-dimension weighted spectrum tensor was established by combine the multiple preset weighting density matrix, and the parallel factor 2 model was introduced to select the exponential factor of weighting function adaptively. Moreover, an minimum spectrum variance evaluation criteria was introduced to improve the estimation precision of separate matrix of mode shape. The identification results of mode shape of a three degree of freedom system show that the proposed method can acquire the best exponential factor of weighting function corresponding to the each mode shape of structure, and the identification precision for mode shape of system is better than traditional frequency domain decomposition method and second-order blind identification method, and it has good identification precision when the measured data length is shorter or large burrs occurs in the frequency spectrum, which provides a new identification method for mode shape of building structures.