Federated Edge AI for Real-Time Environmental Anomaly Detection in IoT Networks

Urban environments face escalating challenges from air pollution and noise, necessitating real-time, fine-grained monitoring for timely intervention and public health protection. Traditional centralized monitoring is often sparse and slow. This paper proposes a novel system for real-time environmental anomaly detection using a network of low-cost IoT sensors (PM2.5 and MEMS microphones) integrated with on-device Machine Learning (ML) models at the edge. To address data privacy and communication overhead, a Federated Learning (FL) framework was implemented, enabling collaborative model training without raw data exchange. This architecture leverages ESP32 microcontrollers for edge inference, significantly reducing latency and cloud dependency. Experimental results demonstrate high anomaly detection accuracy (e.g., 92.1% F1-score for PM2.5 and 92% for noise), achieving inference times under 1 second. Comparison with a simulated centralized approach highlights the FL system's benefits in terms of privacy, communication efficiency (up to 60% data transmission reduction), and robust performance under varying network conditions and sensor faults. This work presents a scalable, privacy-preserving, and efficient solution for smart city environmental monitoring.