The　reduction　of　production　costs　and　the　convenience　of　communication　have　made　the　number　of　smart　meters　in　operation　already　very　large,　resulting　in　later　operation　and　maintenance　work　being　more　difficult.　The　fault　diagnosis　method　for　smart　meters　is　proposed　based　on　stage-wise　additive　modeling　using　a　multi-class　exponential　loss　function　-　classification　and　regression　trees　(SAMMECART).　Firstly,　the　raw　fault　data　of　smart　meter　was　cleaned　by　the　algorithm　of　isolated　forest,　and　the　strong　correlation　index　was　selected　as　the　input　feature　data　using　the　Pearson　correlation　coefficient　method.　Then,　a　multi-classification　model　of　SAMME-CART　was　established　to　classify　five　common　fault　types.　The　experimental　results　show　that　compared　with　other　conventional　classification　algorithms,　the　proposed　method　has　greatly　improved　the　accuracy　of　smart　meter　fault　classification,　and　the　overall　accuracy　rate　reaches　75.72%.　©　2020　IEEE.