A　new　approach　was　proposed　by　combing　Ensemble　Empirical　Mode　Decomposition　(EEMD)　algorithm　and　Back　Propagation　(BP)　neural　network　for　detection　of　gear　through　transmission　noise　analysis.　Then　feature　values　of　the　feature　signals　are　calculated.　The　feature　values　which　have　a　great　difference　for　different　defect　types　are　chosen　to　build　an　eigenvector.　BP　neural　network　is　used　to　train　and　learn　on　the　eigenvector　for　recognition　of　gear　defects　intelligently.　In　this　study,　a　comparative　experiment　has　been　performed　among　normal　gears,　cracked　gears　and　eccentric　gears　with　fifteen　sets　of　different　gears.　Experimental　results　indicate　that　the　proposed　method　can　detect　gear　defect　features　carried　by　the　transmission　noise　effectively.