In order to solve the problems of underdetermined blind source separation for diesel engine crankshaft bearing vibration signal, a determined method based on singular value decomposition by using dynamic clustering and phase space reconstruction was proposed. A quasioptimal embedding dimension and time delay values were first acquired through introducing the concept of generalized embedded window length in order to reconstruct phase space matrix. Then the singular value decomposition was applied to the matrix, the obtained singular values were conducted by using dynamic clustering to determine the optimal reconstruction order, the virtual observation signals were obtained, and hence the underdetermined problems became determined or overdetermined. Finally, the virtual observed signal and the original signal were composed into propagation paths so that the signal containing bearing fault characteristics was extracted by the blind source separation based on the algorithm of adaptive Parafac. The simulation results show that the method can extract the source signal from mixed signals effectively. The diagnosis accuracy increases by 18.4% by applying the method to diesel engine crankshaft bearing fault diagnosis.