Self-Adaptive Sensor Fault Detection in IoT Health Monitoring Using Federated Learning and Lightweight Transformers

Abstract

The spread of Internet of Things (IoT)-based healthcare systems has considerably enhanced real-time patient monitoring, but the performance of such systems largely relies on the integrity and accuracy of sensor data. Heartbeat sensors, although unavoidable, are prone to environmental noise, hardware degradation, and cyber threats, resulting in potential faults that can undermine medical decisions. Addressing these challenges, this paper proposes a novel self-adaptive sensor fault detection framework that leverages Federated Learning (FL) and Lightweight Transformer models. In the proposed system, individual IoT nodes (such as patient monitoring devices equipped with heartbeat sensors) locally train lightweight Transformer-based models to detect sensor faults, while collaboratively updating a global model without transmitting raw health data, thereby preserving privacy and reducing communication overhead. The self-adaptive ability enables continuous learning from new data distributions, making the model robust against changing sensor behaviours and external interference. Comprehensive experiments on real-world heartbeat sensor datasets show that the federated lightweight Transformer framework proposed outperforms classical centralized and standalone approaches with better accuracy, scalability, and robustness. This research offers a promising avenue for the next generation of smart and privacy-aware healthcare monitoring systems to provide reliable and real-time fault detection in IoT-based systems.