A deep federated learning approach for cross-regional railway image analysis

Cross-regional railway data analysis is vital for improving safety and operational efficiency on a global scale. However, the diverse climates, traffic patterns, and operational practices across multiple regions, coupled with strict data privacy regulations, often hinder large-scale centralized training efforts. To overcome these obstacles, we propose a federated deep learning framework that transmits only model parameters after they have been updated within each local node. This approach maintains compliance with privacy requirements while taking advantage of distributed data diversity. Our experiments draw on a dataset of 500,000 images collected from ten geographically distinct railway regions, reflecting both harsh and moderate climates as well as differing infrastructure statuses. The results indicate a 10.2% improvement in foreign object detection accuracy and a 40% reduction in communication overhead compared to single-region models. Such findings highlight the potential of federated learning to unify decentralized data for more effective anomaly detection applications in the railway sector, without compromising confidentiality or incurring excessive training costs. By maintaining shorter communication intervals and employing asynchronous updates, the proposed framework aligns well with real-world railway network structures. Overall, this work underscores the feasibility of privacy-preserving, largescale collaboration in modern railway analytics. Specifically, it can accommodate ongoing expansions in infrastructure.