Federated Learning (FL) clients frequently encounter noisy labeled data due to the costs associated with annotation and the occurrence of malicious attacks. The overparameterized network would suffer from overfitting noisy data resulting in poor generalization ability of parameters. However, many current methods have not considered the causes of this phenomenon and cannot solve noisy labeled data. We propose a Federated Dynamic Aggregation learning scheme based on Parameter Decomposition (FDAPD) for processing noisy labeled data. FDAPD decomposes the model parameters, allowing the parameters to learn clean data and noisy labeled data separately. The client locally trains with time-varying loss functions and dynamic regularization constraints. The central server dynamically aggregates decomposed parameters and reduces the interference of noisy labeled data. Extensive experiments demonstrate that FDAPD can effectively reduce the impact of noisy clients and enhance the generalization ability. Simulations on multiple datasets and models show the superior performance of our method.