The　number　and　proportion　of　Octogenarian　in　China　continue　to　rise,　and　stroke　is　one　of　the　common　diseases　among　elderly　people,　which　seriously　affects　their　healthy　life.　Rehabilitation　training　glove　is　one　of　the　main　rehabilitation　treatment　methods　for　stroke　patients　at　present,　and　its　core　technology　is　surface　electromyography　(sEMG)　gesture　recognition　technology.　Currently,　the　mainstream　EMG　gesture　recognition　model　is　a　convolutional　neural　network　(CNN)　based　on　the　EMG　gesture　recognition　model,　which　has　long　training　and　recognition　time　and　low　training　accuracy　and　cannot　meet　the　current　rehabilitation　needs　of　stroke　patients.　Broad　Learning　System　(BLS)　can　reduce　the　training　and　recognition　time　and　improve　the　training　accuracy,　but　it　cannot　directly　recognize　the　gestures　of　multi-channel　sEMG.　In　this　paper,　we　propose　an　EMG　gesture　recognition　model　based　on　BatchNorm2d　and　incremental　BLS,　which　is　called　BN-BLS,　it　introduces　BatchNorm2d,　the　batch　normalization　layer　of　CNN　network　into　the　incremental　BLS　network　model　to　normalize　the　sEMG　of　each　channel　and　improves　the　processing　performance　of　the　model　for　sEMG.　After　experiments,　compared　with　other　network　models,　the　model　has　short　training　and　recognition　time　and　high　recognition　accuracy,　which　can　be　used　in　rehabilitation　gloves　to　help　stroke　patients　achieve　rehabilitation.　Meanwhile,　the　model　makes　a　breaks-through　of　the　traditional　research　method　of　EMG　gesture　recognition　based　on　deep　learning　and　starts　a　research　idea　of　EMG　gesture　recognition　based　on　broad　learning.　©　2023　IEEE.