In smart cities, wireless communication networks are employed to connect numerous edge devices for the collection and transmission of various types of data. To safe guard data privacy, federated learning is widely adopted in such urban environments. However, densely packed buildings and complex multipath propagation often lead to channel fading, which significantly degrades communication quality and the performance of federated learning. To address this challenge, a reinforcement learning-based adaptive waiting time adjustment algorithm is proposed for wireless federated learning systems. The waiting time is automatically adjusted using reinforcement learning techniques to accommodate dynamically changing channel conditions. Data transmission delays caused by channel fading are mitigated, and the risk of untrustworthy models due to delayed or dropped clients is reduced by optimizing the waiting time. Simulation results on the MNIST and CIFAR-10 datasets demonstrate that, while ensuring overall system performance, the proposed approach maintains high result fidelity and significantly outperforms baseline algorithms across multiple key performance metrics.