MLEDGE — three use cases on one federated learning platform

MLEDGE is a Spanish public-private research project with a deceptively simple test: can a single federated-learning platform serve industries that have almost nothing in common? IMDEA Networks coordinates it; it’s funded by the Spanish Ministry of Economic Affairs and Digital Transformation through EU NextGeneration-EU, running 2023–2025. Acuratio led three of its work streams as the key technological integrator.

Traditional industry. With Inmarepro S.L., we built federated machine-learning models that supervise steam-boiler energy efficiency across four anonymised industrial sites — cosmetics, food, industrial laundry, and pharma. A two-LSTM chain trained on minute-by-minute PLC telemetry predicts the next five minutes of steam output. The federated model lifts R² from 0.44 (local-only) to 0.68 on the data-poorest site, and from 0.25 to 0.59 on the food-production plant — gains a single-site model can’t reach.

Digital economy. With Orange España (37 million active lines), we federated mobility analytics across Madrid and Barcelona for the first half of March in 2020, 2021, and 2024. The pipeline reconstructs anonymised user trajectories, infers home and workplace from geohash CGIs, and detects hospitalisations and self-confinement events. On a representative day in March 2024 in Madrid, the pipeline flagged 7,593 users co-located with a hospital and 1,315 hospitalised for more than four consecutive days — cross-referenced with bed-capacity data to produce an emergency-management heatmap. The work was a finalist at La Chambre’s “Innovation – Large Enterprise” gala.

The FLaaS platform. Underneath both use cases sits the five-layer FLaaS we built from scratch: execution (Docker + Python + Rust core), communications (a WireGuard mesh that punches through corporate NATs without exposing ports), coordination (sees nothing but round signals), governance (per-organisation RBAC), and cost (multi-cloud arbitrage across AWS, GCP, Azure). On top of it, two security primitives developed with IMDEA Networks: FedQV — quadratic-voting aggregation that cuts poisoning-attack success by up to 4× versus FedAvg — and FedRM-RR — repeated-median plus subjective-logic reputation, converging 1.6× to 2.4× faster than baseline robust aggregation while remaining resilient to adversarial clients.

One platform, two very different industries, one set of privacy guarantees. The MLEDGE case studies — steam boilers, telco mobility, and the FLaaS platform — have the full detail.