An　SODDAG-SVM　(structure-optimized　decision　directed　acyclic　graph-support　vector　machine)　multi-classification　action　recognition　method　of　upper　limb　rehabilitation　training　for　the　Brunnstrom　4　~　5　stage　patients　is　proposed　to　solve　the　core　problem　of　action　recognition　of　the　rehabilitation　training　evaluation　method.　First,　the　multi-classification　problem　is　decomposed　into　a　set　of　binary　classification　problems,　support　vector　machine　(SVM)　method　is　used　to　construct　each　binary　classifier,　in　which　the　SVM　kernel　function　parameters　and　feature　subsets　of　each　binary　classifiers　are　optimized　by　genetic　algorithm　and　the　feature　subsets　discrimination　criterion,　respectively.　Then,　the　generalization　errors　of　each　SVM　binary　classifier　are　used　to　measure　the　separable　degree　of　this　class　pair,　and　the　upper　triangulation　matrix　of　generalization　errors　is　built.　Finally,　from　the　root　node,　according　to　the　generalization　error　matrix　of　each　node,　an　SODDAG-SVM　structure　is　constructed　by　choosing　the　SVM　classifier　of　the　most　easily　separated　class　pair　as　each　node.　When　there　are　fewer　instances　to　be　predicted,　a　part　of　the　SODDAG-SVM　structure　passed　by　these　instances　is　directly　built　for　predicting　the　instances.　When　more　instances　need　to　be　predicted,　a　complete　SODDAG-SVM　structure　is　first　constructed　and　then　is　used　to　predict　all　the　instances.　Action　recognition　experiment　is　performed　on　the　upper　limb　routine　rehabilitation　training　samples　of　the　Brunnstrom　4　~　5　stage,　acquired　using　human　body　sensing　technology.　Results　show　that　the　accuracy　reaches　95.49　%　which　is　higher　than　those　of　conventional　decision　directed　acyceic　graph　(DDAG)　and　MaxWins　methods.　It　is　proved　that　the　proposed　method　can　effectively　improve　the　accuracy　of　rehabilitation　training　action　recognition.　Copyright　©　2020　Acta　Automatica　Sinica.　All　rights　reserved.