The　locomotive　traction　motors　convert　electrical　energy　into　mechanical　energy　and　are　the　power　source　of　the　train.　Therefore,　their　working　condition　is　vital　to　the　performance　and　security　of　the　entire　train.　Motor　current　signature　analysis　(MCSA)　has　been　widely　used　in　motor　fault　detection　because　of　its　low　cost　and　accessibility.　The　classic　MCSA　based　on　residual　harmonics　is　only　suitable　for　motor　diagnosis　under　uniform　speed　conditions.　When　the　locomotive　is　running　on　railway　line,　the　speed　cannot　be　kept　constant.　Moreover,　the　reflects　of　mechanical　faults　into　the　stator　current　are　typically　very　subtle　and　easy　to　be　disturbed　in　industrial　field.　To　overcome　these　limitations,　a　novel　motor　fault　characteristic　extraction　method　based　on　the　phase　relations　of　harmonic　component　in　the　stator　current　is　proposed.　The　relationship　between　faults　and　phase　coupling　of　the　harmonics　under　variable-speed　conditions　are　analyzed　and　the　time-varying　quadratic　phase　coupling　(QPC)　models　are　presented.　Then,　the　instantaneous　wavelet　bicoherence　is　employed　to　analyze　the　motor　current　signal　and　the　fault-related　QPC　was　extracted.　Furthermore,　wavelet　bispectrum　entropy　is　proposed　to　describe　the　uniformity　of　QPC　at　the　bifrequency　domain,　which　can　indicate　the　severity　of　failure　from　an　unusual　perspective.　The　proposed　approach　was　effectively　applied　in　the　locomotive　online　operation　test,　and　the　faults　of　the　traction　motor　was　successfully　diagnosed.　IEEE