FL-Transformer: A Privacy-Preserving Approach for Predictive Maintenance of Distributed Industrial Equipment Using Multi-Sensor Edge Intelligence

This paper proposes a privacy-preserving multi-sensor fault prediction system for industrial equipment, integrating federated edge learning and a lightweight Transformer architecture. Addressing the limitations of conventional fault prediction systems—such as inadequate data privacy protection, constrained computational resources, and poor adaptability to heterogeneous devices—our framework combines federated learning (FL) with edge computing to establish a cloud-edge collaborative architecture. On the edge side, a lightweight Transformer model is designed to reduce computational complexity, coupled with a dynamic adaptive sampling mechanism that adjusts sensor sampling frequencies based on real-time device states. On the cloud side, secure aggregation and model management modules enable encrypted parameter uploads from distributed edge devices and global model optimization. Experiments on a self-constructed industrial sensor dataset demonstrate that the proposed system achieves a 3.2% improvement in fault prediction accuracy compared to centralized approaches while preserving data privacy. Additionally, it reduces inference latency by 43.5% and resource consumption by 58.7% on edge devices. Evaluations in both simulated and real-world deployment scenarios validate the system’s feasibility, efficiency, and scalability.