Abstract: Parallel Factor is the multidimensional data processing algorithm, which has the uniqueness of the model decomposition under loose constraints. Combining Local mean decomposition with Parallel Factor, a novel underdetermined blind source separation method is proposed. The production functions (PF) of the original observed signals is obtained by LMD. These PF and original observed signals are combined into new observed signals. Then new signals undergo whitening process and are constructed as the PARAFAC model. Finally the above PARAFAC model is decomposed by trilinear alternating least square (TALS) and the source signals is precisely estimated. The simulation results show that the proposed algorithm can accurately estimate the vibration source signals from the underdetermined mixtures of non-stationary signal. Finally, the proposed method is applied to the multi-source mechanical vibration test, and the experiment results further verify the effectiveness of the approach.