Federated Meta-Learning for Node-Level Failure Detection in Heterogeneous Distributed Systems

Abstract

This paper addresses the challenges of data heterogeneity, privacy protection, and task personalization in server node failure detection within distributed systems. It proposes an intelligent detection algorithm framework that integrates federated learning and meta-learning. The method uses a task-adaptive meta-learning mechanism to build a transferable meta-model. This enables fast adaptation to different node failure patterns and improves generalization. At the same time, a personalized aggregation strategy is introduced to dynamically adjust the model parameter updates based on the local features of each node. This enhances the personalization of local models and improves overall detection accuracy. During model training, data remain stored locally to preserve privacy. This avoids the data leakage risks of traditional centralized learning. The method also enables node-level adaptive optimization while maintaining global collaboration. A series of experiments, including comparative studies, ablation tests, and robustness evaluations, are designed to validate the effectiveness and advantages of the proposed method from multiple perspectives. The results show that compared to mainstream federated learning methods, the proposed model achieves significant improvements in Accuracy, Precision, and Recall. It performs especially well under complex scenarios such as Non-IID data and inactive nodes. These findings demonstrate the method's strong stability, adaptability, and practical potential for large-scale distributed environments.