In large rotating machinery fault diagnosis, the number of fault sources is varied, so it is difficult to estimate the source number accurately and separate the fault sources effectively. Exhibition four-order cumulant matrix is adopted to estimate the source number of machinary fault diagnosis and according to the relation of source number and sensor number, corresponding source separation algorithm is chosen to separate blind source correctly. The experiment indicates that the estimating algorithm can estimate out sensor number effectively in cluding under-determined, the algorithm of adaptive blind source separation can realize the fault diagnosis effectively including over-determined, under-dete rmined and determined.