Federated learning （FL）， as a privacy-preserving distributed learning framework， has solved the data silo problem effectively. However， the heterogeneous characteristics of Internet of Things （IoT） terminal devices in real-world scenarios， particularly device performance variations and Non-Independent and Identically Distributed （Non-IID） data properties， will cause degradation of the model performance and convergence speed. To address these challenges， a federated split learning optimization method under edge heterogeneity named FedCRS （Federated Cluster-based Round Splitting） was proposed. Firstly， an adaptive clustering strategy based on device performance was developed to cluster the clients dynamically， and assign customized sub-models for the clusters to balance computational loads， thereby solving straggler effect caused by device heterogeneity. Then， a ring-topology cyclic model transfer mechanism was created， where local model fusion of intra-cluster clients was implemented to alleviate client drift caused by data heterogeneity while enhancing global model robustness and generalization capability significantly. Experimental results on three label-heterogeneous datasets （FMNIST， CIFAR-10， and CIFAR-100） demonstrate that compared with five baseline methods （FedAvg （Federated Averaging）， FedProx， MOON （MOdel-cONtrastive learning）， SplitFed， and SplitMix （Split Mixing） ）， FedCRS has the best accuracy on all the datasets， and has the accuracy improvements of at least 8.7， 11.1， and 2.1 percentage points， respectively； and on FMINST， CIFAR-10 datasets， FedCRS has the convergence acceleration reached 78.1%and13.2%， respectively. It can be seen that FedCRS is effective in optimizing model accuracy and convergence speed under edge heterogeneous environments， indicating good practical application prospects.