Objective.　The　surface　electromyography　(EMG)　signal　reflects　the　user＇s　intended　actions　and　has　become　the　important　signal　source　for　human-computer　interaction.　However,　classification　models　trained　on　EMG　signals　from　the　same　day　cannot　be　applied　for　different　days　due　to　the　time-varying　characteristics　of　the　EMG　signal　and　the　influence　of　electrodes　shift　caused　by　device　wearing　for　different　days,　which　hinders　the　application　of　commercial　prosthetics.　This　type　of　gesture　recognition　for　different　days　is　usually　referred　to　as　long-term　gesture　recognition.　Approach.　To　address　this　issue,　we　propose　a　long-term　gesture　recognition　method　by　optimizing　feature　extraction,　dimensionality　reduction,　and　classification　model　calibration　in　EMG　signal　recognition.　Our　method　extracts　differential　common　spatial　patterns　features　and　then　conduct　dimensionality　reduction　with　non-negative　matrix　factorization,　effectively　reducing　the　influence　of　the　non-stationarity　of　the　EMG　signals.　Based　on　clustering　and　classification　self-training　scheme,　we　select　samples　with　high　confidence　from　unlabeled　samples　to　adaptively　updates　the　model　before　daily　formal　use.　Main　results.　We　verify　the　feasibility　of　our　method　on　a　dataset　consisting　of　30　d　of　gesture　data.　The　proposed　gesture　recognition　scheme　achieves　accuracy　over　90%,　similar　to　the　performance　of　daily　calibration　with　labeled　data.　However,　our　method　needs　only　one　repetition　of　unlabeled　gestures　samples　to　update　the　classification　model　before　daily　formal　use.　Significance.　From　the　results　we　can　conclude　that　the　proposed　method　can　not　only　ensure　superior　performance,　but　also　greatly　facilitate　the　daily　use,　which　is　especially　suitable　for　long-term　application.