Abstract: There exist various vibration sources in a hydropower house, and how to accurately identify them is crucial to assessing the dynamic safety of the house. In this paper, we describe a method for blind separation of the vibration sources in a large scale underground hydropower house that combines the techniques of filtering denoising, estimating the number of vibration sources, and the joint approximate diagonalization of eigen-matrix (JADE). First, a filtering denoising method is used to de-noise all the multi-dimensional signals observed. Then, the correlation matrix of the signals is solved, and the number of vibration sources is estimated using the dominant eigenvalue and Bayesian information criterion . Finally, the signals are pre-whitened, and separated using a JADE method. This blind separation method is verified through analog signal processing. When applied to the vibration signals in a hydropower house, it accurately separates vortex belt, unit rotation, and volute flow uniformity. Thus it is an effective method for exploring the characteristics of vibration sources in hydropower houses.