Federated　learning　has　been　successfully　used　in　the　Internet　of　Things　(IoT)　scenarios.　However,　the　data　among　IoT　devices　are　usually　Non-IID　(Identically　and　Independently　Distributed),　this　data　heterogeneity　will　severely　affect　the　convergence　rate　of　the　global　model　and　even　result　in　a　significant　decrease　in　model　accuracy.　To　address　this　problem,　this　paper　proposes　a　federation　adaptive　aggregation　algorithm　based　on　the　representation　capabilities　under　feature　alignment.　Prior　to　federation　training,　the　server　computes　initial　feature　anchors　using　the　global　model　with　Xavier　uniform　initialization,　thus　making　it　more　responsive　to　data　features.　During　client　training,　the　feature　alignment　method　updates　all　client　models　within　the　uniform　feature　space　using　enhanced　feature　anchors,　thus　reducing　the　impact　of　data　heterogeneity　between　clients.　A　weight　scaling　factor　related　to　the　representation　capability　is　introduced　during　the　server　federated　aggregation　process　to　adaptively　aggregate　local　model　parameters　in　the　presence　of　data　heterogeneity.　Extensive　experiments　were　conducted　on　popular　image　datasets,　namely　MNIST,　Fashion-MNIST,　CIFAR-10,　and　CIFAR100.　The　experimental　results　demonstrate　that　our　method　significantly　improves　the　convergence　speed　of　the　global　model　and　the　model　performance,　and　at　the　same　time　effectively　mitigates　the　detrimental　effects　of　data　heterogeneity.