A　new　method　of　mixed　fault　diagnosis　was　proposed　for　power　electronic　circuit　based　on　auto-　regressive　moving　average　(ARMA)　and　discrete　hidden　Markov　model　(DHMM).　Firstly,　the　fault　circuit　sampling　data　was　processed　by　mean　normalization　method.　Afterwards,　characteristic　quantities　of　circuit　fault　information　were　analyzed　and　extracted　by　using　ARMA　bispectrum　analysis,　and　then　it　was　delt　with　vector　quantization.　Finally　the　discrete　hidden　Markov　models　were　utilized　to　design　the　fault　classifier　of　power　electronic　circuits.　The　method　was　applied　to　fault　diagnosis　of　the　SS8　electric　locomotive　main　converter.　The　results　show　that　the　proposed　method　has　high　diagnostic　accuracy　and　good　ability　to　resistance　the　noise.　The　diagnosis　correct　rate　of　the　proposed　method　is　100%　in　the　case　of　no　noise　or　adding　5%　noise.　When　adding　10%　noise,　its　diagnosis　correct　rate　is　16.11%　and　23.79%　higher　respectively　than　that　of　DHMM　and　GA-BP　neural　network　method.　The　diagnosis　method　is　useful　in　the　engineering.　©　2010　Chin.　Soc.　for　Elec.　Eng.