Low-Power Microcontroller-Based Decentralized Distributed Sequential Neural Network for Anomaly Detection in Wireless Sensor Networks
Yernazar Bolat, Nasim Ferdosian, Iain D. Murray · 2025
As IoT applications grow, low-power Microcontroller Units (MCUs) are increasingly used for energy-efficient, local data processing. However, deploying complex Deep Neural Networks (DNNs) on such constrained devices remains challenging due to limited memory, computational power, and energy. Traditional solutions rely on cloud offloading or model compression, which can introduce bandwidth overhead, privacy risks, and accuracy loss. This paper presents a Decentralized Distributed Sequential Neural Network (DDSNN) approach to address these challenges by partitioning a DNN across multiple MCUs, enabling layer-wise, sequential inference without relying on a central coordinator. Unlike other distributed works that are dependent on a central coordinator, making them cluster-based computing, we are the first to implement fully Decentralized Distributed Deep Neural Network inference in Wireless Sensor Networks. To validate DDSNN, we deploy it in a real-world predictive maintenance scenario, where vibration data is analyzed in real-time to predict abnormal operating conditions of an industrial pump. We show that, by using our approach on hardware commonly employed in TinyML applications (under 512 KB of RAM), we achieve 99.01% accuracy, demonstrating the practical feasibility of DDSNN for TinyML deployments in WSNs.