Federated　learning　has　been　successfully　used　in　the　Internet　of　Things　(IoTs)　because　it　breaks　data　silos　and　protects　data　privacy　and　security　by　training　shared　global　models　through　multi-client　collaboration.　However,　the　data　distribution　heterogeneity　across　IoT　devices　will　severely　slow　down　the　convergence　of　the　global　model　and　even　result　in　a　significant　decrease　in　model　accuracy.　To　address　this　issue,　a　federated　adaptive　aggregation　algorithm　was　developed　based　on　Coefficient　of　Variation.　The　proposed　algorithm　adaptively　adjusts　the　aggregation　weight　factor　of　each　local　client　model　according　to　the　difference　in　the　coefficient　of　variation　obtained　before　and　after　local　model　training.　And　the　proportion　of　the　current　global　model　in　the　aggregation　process　is　also　dynamically　adjusted　in　the　middle　and　later　stages.　Extensive　experiments　demonstrate　that　the　proposed　method　improves　the　accuracy　of　the　global　model　and　is　also　robust　during　the　federated　training　process.