In UAV-assisted edge computing, centralized resource allocation can lead to privacy leakage, and issues such as data heterogeneity, sudden user distribution changes, and strategy delays or oscillations may result in resource wastage and low user satisfaction. To address these problems, a resource allocation algorithm based on personalized federated deep reinforcement learning (P-FDRL-RA) is proposed in this paper. The algorithm integrates federated learning (FL) with deep reinforcement learning (DRL), using a “distributed training - centralized aggregation” framework to optimize global performance. Additionally, a personalized model is introduced to allow each drone to adapt to local environmental differences. In the personalized model, an adaptive learning rate adjustment and a regularization optimizer (AdamW) are employed, with the global model serving as a regularization term to adjust the training process. This ensures that each UAV’s DDQN training update is better suited to its own data, effectively overcoming the gradient conflict problem caused by learning rate adaptation in federated learning. Experimental results demonstrate that the P-FDRL-RA algorithm significantly enhances network throughput and user satisfaction when compared to other similar algorithms.