To　achieve　compound　faults　diagnosis　with　single　channel　signal,　a　three-dimensional　geometric　features-based　sparse　component　analysis　(TGF-SCA)　method　is　proposed.　Intrinsic　characteristic-scale　decomposition　(ICD)　is　used　to　decompose　the　single　channel　of　mixed　signal　into　three　channels.　Then,　the　three-dimensional　potential　function　(TPF)　is　constructed　based　on　the　three-dimensional　geometric　features　to　estimate　matrix.　In　addition,　an　energy　factor　(EF)　is　introduced　to　improve　the　computational　efficiency　in　the　process.　Ultimately,　the　minimal　l(1)　norm　algorithm　is　used　to　obtain　the　separated　signal　based　on　the　estimated　matrix.　Experimental　analysis　results　for　roller　bearing　show　that　the　fault　feature　frequencies　of　bearings　acquired　using　the　proposed　approach　are　evidently　close　to　the　theoretical　values.　For　example,　when　the　rotating　speed　is　900　rpm,　the　feature　frequency　60.27　Hz　is　very　similar　to　the　theoretical　calculation　of　ball　pass　frequency　of　the　outer　race　(BPFO)　60.5　Hz　and　the　feature　frequency　74.01　Hz　is　close　to　the　theoretical　calculation　of　the　ball　pass　frequency　of　the　roller　(BPFR)　74.4　Hz.　Compared　with　the　ICA　method,　the　SCA　method　based　on　Fuzzy　C-means　algorithm　(FCM)　and　the　SCA　method　based　on　K-means　algorithm,　the　experimental　verification　results　indicate　that　the　TGF-SCA　method　can　separate　the　source　signal,　extract　the　fault　features　and　realize　compound　faults　diagnosis　for　roller　bearing.　(C)　2018　Elsevier　Ltd.　All　rights　reserved.