An IoT-based privacy-preserving federated learning framework for proactive machine maintenance prediction in Industry 5.0

Abstract

This study aims to propose an Internet of Things (IoT)-based, privacy-preserving Federated Learning (FL) framework for predicting machine failures in Industry 5.0. In this study, we provide a reference design in which sensor data are collected from industrial machines spread across multiple buildings. Each machine is equipped with IoT sensors that measure vibration, humidity, temperature, and pressure. Our methodology compares ten AI models across two different approaches: Deep Learning (DL) models (Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), Autoencoders, Convolutional Neural Network (CNN), Transformers, and EfficientNet1D) and Federated Learning models (Channel-Separated CNN-FL, Hierarchical FL, Adaptive FL, Ensemble FL), all of which are evaluated using standard accuracy metrics. Federated Learning models perform better than Deep Learning methods, both in terms of accuracy and practical use. Hierarchical FL achieved the highest accuracy, reaching 98.44%, with a precision of 99.24%. EfficientNet1D showed the best recall, at 87.58%. These results confirm that FL models provide data privacy through decentralized training while preserving accuracy, making them ideal for industrial and enterprise-scale applications. Some divergence was observed in FL, and Deep Learning models exhibit high computational complexity. The findings enable enterprises and industries to transition from reactive to predictive maintenance, which would help them reduce unplanned downtime and operational costs. This framework is consistent with Industry 5.0, wherein AI handles monitoring and humans concentrate on complex operations. This study integrates Deep Learning and Federated Learning methodologies for predictive maintenance within the context of Industry 5.0, specifically addressing the frequently neglected concerns of data privacy and decentralized operational settings.