A Pr edictive Maintenance Feder ated Lear ning Fr amewor k for Machine Fault Detection Based on Vibr ation Data

Predictive maintenance is essential in the manufacturing industry, as it helps reduce maintenance costs and improve operational efficiency. In the IoT era, predictive maintenance is increasingly automated using deep learning methods, leveraging the abundant availability of sensor data. However, the widely used centralized learning approach presents significant data security concerns, as company data is highly valuable and must be protected. In this paper, we propose a Federated Learning (FL) framework for predictive maintenance, utilizing vibration data collected from edge devices. Additionally, we introduce a custom model aggregation strategy, Fine-Tuning FedAvg (FT-FedAvg), which enhances model performance by incorporating a fine-tuning phase during the model aggregation process. Our framework is designed for real-world deployment, as it is optimized for resource-constrained edge devices and supports real-time monitoring in industrial environments. Experimental results demonstrate that our FL framework achieves comparable performance to centralized learning while preserving data privacy, making it a practical and secure solution for predictive maintenance in real-world manufacturing scenarios.