Federated Learning for Fault Detection in Edge-Based IoT Systems

Edge-based Internet of Things (IoT) deployments increasingly support safety- and cost-critical operations, requiring reliable fault detection under strict bandwidth, latency, and data-governance constraints. Centralized learning pipelines are often impractical because high-rate sensor telemetry is expensive to transmit and may be restricted by privacy or proprietary policies. Federated learning (FL) addresses these limitations by enabling collaborative model training across distributed edge clients while keeping raw data local. However, fault detection in edge-IoT introduces strong statistical and systems heterogeneity, including non-IID operating conditions, label imbalance, intermittent participation, and client drift, which can degrade standard FL approaches. This study investigates FL for fault detection using a unified benchmark suite spanning vibration fault classification (CWRU and Paderborn), acoustic anomaly detection (MIMII), and multivariate degradation/anomaly detection (NASA C-MAPSS). We evaluate FedAvg against heterogeneity-aware optimizers (FedProx and SCAFFOLD) under non-IID client partitions and hierarchical edge-gateway-cloud training. Results show that federated methods consistently outperform local-only training and approach centralized upper bounds while preserving data locality. Robust FL methods provide statistically significant improvements over FedAvg under severe heterogeneity, with the largest gains on datasets exhibiting strong domain and trajectory skew. Communication analysis highlights trade-offs between robustness and overhead and shows that compression can substantially reduce bandwidth with modest impact on convergence. These findings support FL as a practical pathway for scalable, privacy-preserving fault detection in edge-based IoT systems and motivate future work on drift-aware continual federated learning, personalization, and trustworthy aggregation.