<p>Federated learning (FL) collaboratively generates a single global model from distributed clients without compromising data privacy. However, data heterogeneity has a significant impact on model training in federated learning. Moreover, in the presence of heterogeneity, the global model exhibits poor generalization across clients, often containing information that individual clients do not expect. To address this issue, we propose a personalized federated learning algorithm called personalized federated learning with adaptive optimization of local model (FedAom), which improves the local models on clients that participate in local training. This is achieved by retaining the locally trained models obtained in each communication round to form personalized models, and comparing them with the new global model passed from the server in the next round of training, a more suitable initial model for local data training is generated for the client, thereby improving the effectiveness of local training. To evaluate the effectiveness of the algorithm, we conducted extensive experiments on three benchmark datasets, demonstrating that FedAom achieves a testing accuracy improvement of approximately 3.16% over baseline algorithms. Furthermore, we applied the FedAom algorithm to other federated learning algorithms, resulting in a testing accuracy improvement of around 3.03%.</p>