According　to　the　problems　that　the　vibration　features　of　bearing　faults　were　hard　to　separate　and　recognize　in　strong　vibration　source　inhibition,　a　diagnosis　method　was　established　based　on　source　number　estimation　and　FDBD　algorithm.　Wavelet　packet　decomposition　was　used　to　divide　the　signals　into　multiple　sub　band　signals,　and　SVD　was　selected　to　estimate　the　signal　source　numbers　in　underdetermined　conditions.　The　multiple　dimension　signals　were　constructed　based　on　the　source　number　estimation.　The　FDBD　algorithm,　which　included　STFT,　fast-ICA　in　complex　domain,　relevance　ranking　and　inverse　STFT,　was　finally　applied　on　fault　feature　separation　and　extraction.　The　effectiveness　of　the　method　was　validated　in　fault　feature　separation　and　weak　feature　recognition　by　the　simulation　signals　and　experimental　data　of　rolling　bearing　faults.　©　2017,　Chinese　Mechanical　Engineering　Society.　All　right　reserved.