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

Area of Science:

• Signal Processing
• Machine Learning
• Internet of Things (IoT)

Background:

• IoT devices generate vast amounts of time-series data requiring efficient processing.
• Real-time signal analysis is critical for applications like predictive maintenance and environmental sensing.
• Edge computing offers solutions to reduce latency and communication overhead in IoT networks.

Purpose of the Study:

• To propose a lightweight framework for edge-based anomaly detection in IoT sensor networks.
• To combine classical signal processing with edge machine learning for efficient data analysis.
• To evaluate the performance and resource efficiency of different models for anomaly detection.

Main Methods:

• Feature extraction using Fourier and Wavelet transforms.
• Deployment of machine learning models (Shallow Neural Network, Quantized TinyML, Decision Trees) on edge devices.
• Testing with synthetic vibration, acoustic, and environmental time-series datasets.

Main Results:

• Shallow Neural Network achieved high detection performance (F1-score ≈ 0.94).
• Quantized TinyML model provided a good balance of performance (F1-score ≈ 0.92) with significantly reduced memory and energy usage.
• Decision Trees offered low latency suitable for highly constrained devices.

Conclusions:

• The proposed framework enables accurate and resource-efficient anomaly detection directly on IoT edge devices.
• Edge-based processing is feasible for large-scale IoT sensor networks, improving operational efficiency.
• The approach demonstrates robustness against common data imperfections and variations.