The　magnitude　determination　of　earthquake　is　one　of　the　most　important　and　challenging　parts　in　Early　Earthquake　Warning　(EEW)　system.　In　this　paper,　the　method　of　determination　of　earthquake　magnitude　based　on　convolutional　neural　network　(CNN)　is　proposed.　This　method　transforms　magnitude　determination　problem　into　a　classification　problem,　by　dividing　earthquake　magnitudes　into　20　different　categories　which　are　greater　than　2.0　(ML＞2.0).　In　this　paper,　a　total　of　1928　earthquakes　in　Fujian,　Taiwan　Strait　and　Taiwan　area　recorded　by　Fujian　Seismic　Network　from　2012　to　2019　were　collected　as　research　data;　14644　three　component　seismic　records　were　obtained　by　station　record　interception,　data　tagging　and　quality　screening　along　with　other　pre-processing　procedures.　A　convolutional　neural　network　(CNN)　model　for　magnitude　prediction　was　constructed　by　inserting　three-second　data　records.　The　model　was　trained　with　the　earthquake　events　from　2012-2018　and　tested　with　the　earthquake　events　in　2019.　The　results　showed　that　85.6%　of　the　magnitude　deviation　of　a　single　station　can　be　controlled　within　0.3,　and　91.8%　of　the　average　magnitude　deviation　of　the　first　three　stations　can　be　controlled　within　0.3.　Those　cases　with　relatively　large　deviation　are　mainly　due　to　the　lack　of　historical　samples.　Compared　with　the　traditional　methods,　the　magnitude　determined　by　the　CNN　model　is　more　stable　and　reliable,　which　can　provide　a　new　technical　method　for　solving　the　challenging　problem　of　EEW　magnitude　determination.　©　2021,　Science　Press.　All　right　reserved.