Research on Autonomous Path Planning and Control Strategy of UAV Based on Multi-Objective Optimization

Wireless Sensor Networks (WSNs) often operate in environments where data distribution changes over time, leading to performance loss in trained models. This issue, known as concept drift, is not well addressed by standard Federated Learning (FL) frameworks, which typically assume that the data remains unchanged. To address this limitation, we propose a lightweight Federated Continual Learning (FCL) framework suitable for WSNs with limited resources. The framework includes three main components: a simple drift detection method, a selective memory update strategy, and an adaptive learning rate adjustment. These components work together to maintain model accuracy as the data evolves. Experiments on a synthetic WSN dataset show that the proposed method reduces pitch prediction error by 12.3% and roll prediction error by 9.8% compared to standard FL. It also achieves a 17.5% reduction in communication cost. These results demonstrate that the proposed FCL framework can improve long-term performance in WSNs under non-stationary conditions and is suitable for real-world applications where computing and communication resources are limited.