Global manufacturers want one vision or predictive model that learns from every plant—and legal, customer, and OT teams refuse to centralize the raw data. Federated learning addresses that tension: train locally, share weight updates or gradients, aggregate a global model, and push improvements back—without exporting frames, recipes, or personnel records.
In 2026 the pattern moved beyond research papers into multi-site quality and maintenance pilots, especially where GDPR, defense, or OEM IP rules block a single training lake.
What federation is good for
- Cross-site generalization — Rare defects at one plant teach others.
- Data residency — Keep OT imagery and traces inside the fence.
- Partner networks — Suppliers collaborate on models without opening process IP.
What still fails in production
Non-IID data across plants breaks naive aggregation. Communication overhead and stragglers slow rounds. Poisoned or low-quality local updates can degrade the global model without strong governance. And federation does not remove the need for local validation before a site promotes weights to a live cell.
Security of the aggregation server and update channels matters as much as model accuracy—gradient leakage and model theft are real OT concerns.
What to watch next
- Named multi-plant programs publishing scrap or downtime gains versus single-site baselines.
- Standard MLOps patterns for signed federated rounds into MES/QMS-controlled promotion.
- Whether industry consortia host neutral aggregators—or each OEM builds a private federation.
Edge inference runs the model; synthetic data fills gaps. Federated learning is the network chapter: improve together while the raw plant stays home.