Since　roller　bearing　is　one　of　the　most　vulnerable　components,　bearing　faults　usually　occur　in　an　unprepared　situation　with　multiple　faults,　and　the　quantity　of　sensors　is　limited　in　the　real-time　working　environment,　resulting　in　an　underdetermined　blind　source　separation　(UBSS)　problem　to　extract　the　fault　features.　Because　the　collected　signals　are　usually　not　independent　and　not　sparse　enough,　traditional　methods　of　separating　signals　cannot　perform　well.　In　this　paper,　an　optimized　intrinsic　characteristic-scale　decomposition　(OICD)　method　is　proposed　to　solve　the　underdetermined　problem.　Meanwhile,　the　constraint　error　factor　is　introduced　to　overcome　the　drawback　that　the　ideal　ending　condition　of　ICD　is　not　proper　for　the　vibration　signal　of　bearing.　In　addition,　given　that　non-negative　matrix　factorization　(NMF)　is　not　limited　by　the　source　signal　independence　and　sparsity,　an　improved　UBSS　model　is　constructed,　and　the　PCs　are　used　as　the　input　matrix　of　local　NMF　to　obtain　the　separation　signal.　Ultimately,　envelope　analysis　is　utilized　to　detect　the　source　signal　feature.　Both　simulated　and　experimental　vibration　signals　are　used　to　verify　the　effectiveness　of　the　proposed　approach.　Besides,　the　traditional　method　is　juxtaposed　with　the　suggested　method.　The　results　indicate　that　the　proposed　method　is　effective　in　dealing　with　the　compound　faults　separation　of　the　rotating　machinery.