Towards IoT Rejuvenation: From Machine to Federated Learning Based Regression to Detect Sensor Ageing in Cloud–Edge Continuum

The adoption of low-cost IoT sensors at scale is often limited by hardware and software ageing, which degrades measurement accuracy over time and threatens the reliability of smart environments such as smart cities and smart hospitals. IoT rejuvenation is a proactive paradigm that combines measurement, calibration, and optimisation algorithms to keep sensors accurate throughout their lifecycle. Recent studies have shown that machine learning (ML) regression models can effectively detect the ageing of sensors, providing a viable alternative to analytically derived transfer functions and enabling softwarebased recalibration on the Cloud-Edge continuum. Building on this line of work, this paper introduces a federated learning (FL) approach in which multiple smart lighting poles collaboratively train regression models without sharing raw data, thus enhancing scalability and privacy while preserving predictive performance. Using a Flower-based FL testbed, we compare centralised and federated training (considering IID and non-IID client partitions) of three regression families (SGD, MLP, and an RBF-kernelapproximated SVR) on an experimentally collected sensor ageing dataset, and we show that non-IID, day-wise client partitions can achieve comparable or slightly improved robustness with respect to IID settings, while maintaining inference latencies compatible with edge deployment.