Lightweight Signal Processing and Edge AI for Real-Time Anomaly Detection in IoT Sensor Networks

The proliferation of IoT devices has created vast sensor networks that generate continuous time-series data. Efficient and real-time processing of these signals is crucial for applications such as predictive maintenance, healthcare monitoring, and environmental sensing. This paper proposes a lightweight framework that combines classical signal processing techniques (Fourier and Wavelet-based feature extraction) with edge-deployed machine learning models for anomaly detection. By performing feature extraction and classification locally, the approach reduces communication overhead, minimizes latency, and improves energy efficiency in IoT nodes. Experiments with synthetic vibration, acoustic, and environmental datasets showed that the proposed Shallow Neural Network achieved the highest detection performance (F1-score ≈ 0.94), while a Quantized TinyML model offered a favorable trade-off (F1-score ≈ 0.92) with a 3× reduction in memory footprint and 60% lower energy consumption. Decision Trees remained competitive for ultra-constrained devices, providing sub-millisecond latency with limited recall. Additional analyses confirmed robustness against noise, missing data, and variations in anomaly characteristics, while ablation studies highlighted the contributions of each pipeline component. These results demonstrate the feasibility of accurate, resource-efficient anomaly detection at the edge, paving the way for practical deployment in large-scale IoT sensor networks.