In　order　to　solve　the　problem　of　small-samples　and　incipient　fault　prognosis,　a　novel　identification　approach　based　on　physical　model　is　presented　for　automatic　diagnosis　of　defective　rolling　element　bearings.　The　major　advantage　of　this　method　is　that　its　training　can　be　performed　using　simulation　data.　Prediction　of　the　vibration　response　due　to　defect　requires　an　accurate　model.　Multi-body　dynamics　of　rolling　element　bearing　are　developed　according　to　the　vibration　transmission　path　combining　with　dynamics　contact　mechanism　of　interface.　For　the　purpose　of　extracting　the　feature　of　weak　impact　component,　a　new　detecting　method　based　on　Blind　deconvolution　and　Kurtosis-Laplace　wavelet　is　proposed.　The　simulation　and　the　detection　of　engineering　faint　impact　signal　results　demonstrate　that　this　method　is　highly　effective　in　noise　reduction　and　fault　feature　extraction.　Then,　through　translating　the　inverse　problem　into　geometric　distance　matching,　the　defects　can　be　predicted.　Finally,　experimental　data　is　used　to　verify　the　feasibility　and　reliability　of　current　method.