When　binary　tree　SVM　is　used　for　multi-class　fault　diagnosis,　inner-class　distance　or　between-class　distance　is　always　used　to　decide　the　classification　hierarchy,　but　these　methods　cannot　take　the　comprehensive　separability　information　between　classes　into　account,　which　leads　to　decrease　the　accuracy　of　fault　diagnosis　easily,　so　an　improved　binary　tree　SVM　method　is　proposed.　Combining　the　separability　of　inner-class　with　the　separability　of　between-class,　a　measurement　formula　is　built,　which　is　based　on　a　principle,　that　is　the　same　class　is　relatively　clustered　and　the　different　classes　have　a　relatively　far　distance　is　easier　to　classify.　Then　according　to　it,　the　classification　hierarchy　is　decided.　In　the　end,　the　new　method　is　applied　to　fault　diagnosis　of　Tennessee　Eastman　(TE)　process,　the　experimental　results　show　it　has　an　excellent　integrated　performance　in　comparison　to　other　methods　based　on　SVM.　©　2015　IEEE.