Federated learning for network anomaly detection and visual quality control in shared production

Federated Learning (FL) addresses data sovereignty and heterogeneity in industrial cyber-physical systems by training models across sites without centralizing raw data. We report on two real-world use cases: anomaly detection in operational technology (OT) network communication, and visual inspection for quality control in a shared production scenario across several smart factories at different locations. Both settings feature non-IID data distributions, domain shift, strict privacy/IP constraints, and edge-cloud resource limitations. We benchmark local, centralized, and federated training with domain-specific architectures. Results indicate that FL improves precision-recall behavior for rare events in OT networks at identical ROC-AUC compared to baselines and approaches model quality of the globally learned model in visual inspection with appropriate round-epoch scheduling, while local models overfit site-specific data. We discuss design guidelines, benefits for shared production, and how FL could contribute to scalability and data and energy efficiency in a dynamic edge–cloud continuum.