A Deep Neural Framework for Fault Detection in IoT-Based Sensor Networks

Simona Cicero, Massimo Guarascio, Antonio Guerrieri, Simone Mungari, Andrea Vinci

Abstract: Effective disaster management and early fault detection are crucial for preserving safety and operational efficiency in smart workplaces and buildings. These environments rely on networks of interconnected IoT sensors to monitor parameters such as temperature, humidity, occupancy, and energy usage. Promptly identifying and responding to anomalies or faults using data from these sensors is vital to prevent minor issues from escalating into major disruptions or safety hazards. Recently, Artificial Intelligence techniques have proven to be effective solutions in this scenario, offering sophisticated tools for analyzing complex data patterns. Nevertheless, developing reliable models needs to face a number of important challenges, including the lack of labeled data, unbalanced class distributions, and the need for zero-shot anomaly detection. Moreover, models must be lightweight to run efficiently on IoT devices with limited computational resources and energy supply. To address these challenges, we propose an unsupervised deep learning framework for fault detection in IoT-based sensor networks. Our solution employs a Sparse U-Net architecture, an autoencoder equipped with skip connections and sparse layers to enhance learning and robustness to noise. It uses online training with a sliding-window approach to adapt to evolutions in data distribution without requiring prior knowledge of anomalies. Extensive experimentation on a real test case demonstrates the framework’s effectiveness and efficiency in ensuring the safety of IoT-based smart workplaces.