As　a　safe　and　efficient　distributed　machine　learning　technology,　federated　learning　has　been　widely　used　in　all　walks　of　life.　However,　various　problems　in　the　traditional　FedAvg　algorithm　have　brought　a　lot　of　inconvenience　to　the　implementation　of　federated　learning　in　related　industries.　Therefore,　this　paper　addresses　related　problems　by　proposing　a　novel　federated　learning　algorithm　(N-FedAvg)　based　on　FedAvg.　N-FedAvg　selects　clients　in　sequence　before　each　epoch,　reducing　the　randomness　when　selecting　clients,　so　that　all　clients＇　data　can　have　the　opportunity　to　participate　in　federated　learning,　and　prevent　local　data　of　some　clients　from　participating　in　the　federation　with　a　low　probability.　Also,　when　using　gradient　descent　to　optimize　the　objective　function　and　approaching　the　global　minimum　of　the　loss　function　value,　the　learning　rate　should　be　made　smaller　to　get　the　model　as　close　as　possible　to　this　point,　and　cosine　annealing　can　be　achieved　by　the　cosine　function　and　produces　good　results.　Finally,　the　sparsity　strategy　is　essentially　a　model　compression　method,　which　not　only　transmits　a　small　number　of　parameters,　but　also　reduces　the　network　bandwidth　between　the　server　and　the　clients,　and　can　also　prevent　the　leakage　of　global　model　parameters.　Through　comparative　experimental　analysis,　this　paper　finds　that　the　N-FedAvg　algorithm　proposed　by　us　is　1.34%　more　accurate　than　the　traditional　FedAvg　algorithm　on　the　CIFAR-10　dataset,　and　the　loss　function　value　is　2.77%　lower.　©　2022　IEEE.