Ekkono Synthesis: Federated Learning for the Industrial IoT

Project Information

Ekkono FLIIoT

Grant agreement ID: 190171211

DOI

10.3030/190171211

Project closed

EC signature date 20 April 2023

Start date 1 May 2023

End date 31 December 2025

Funded under

• The European Innovation Council (EIC)

Total cost

€ 3 571 250,00

EU contribution

€ 2 499 875,00

Coordinated by

EKKONO SOLUTIONS AB
Sweden

Periodic Reporting for period 2 - Ekkono FLIIoT (Ekkono Synthesis: Federated Learning for the Industrial IoT)

Reporting period: 2024-07-01 to 2025-12-31

Summary of the context and overall objectives of the project

Context
Across Europe, connected equipment plays an increasingly important role in ensuring efficient, reliable and sustainable operation in both industrial and field environments. These devices often operate in customer contexts where operational data is confidential, OEM access is restricted, and streaming all data to the cloud is not practical or desirable. As a result, traditional centralised machine-learning approaches are difficult to apply in real operational settings. Increasingly, these devices rely on machine-learning models that must be monitored and improved over time to remain reliable. At the same time, Europe is pursuing improvements in efficiency, sustainability and resilience in line with initiatives such as the Green Deal and Industry 5.0 creating a strong need for practical, decentralised AI solutions.

Overall Objectives
The Ekkono FLIIoT project set out to develop Synthesis, a system for managing machine-learning models directly on distributed devices without requiring access to raw data. Synthesis was designed to keep models up to date in the field by enabling confidentiality-preserving learning through federated and model-centric techniques, automatically detecting anomalies, under-performing devices and model degradation, providing meaningful fleet-level insights while keeping data at the edge, and delivering a scalable, intuitive interface capable of handling thousands of models in real time. These objectives support more efficient, resilient and sustainable operation across diverse deployments.

Project Pathway to Impact
Synthesis progressed from core model-centric and federated learning techniques to a scalable platform capable of operating across varied real-world contexts. The completed system was validated in two major sectors: manufacturing and industrial equipment and energy and thermal systems. It demonstrated its capability to support improved operational reliability and efficiency, with promising indications of consistent long-term performance. By enabling these capabilities without transferring raw operational data, Synthesis addresses practical constraints faced by OEMs and operators. The resulting impact includes reduced downtime, better resource use and more informed operational decision-making, contributing to more resilient, efficient and sustainable systems in line with Europe’s broader transition goals.

Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far

During the project, the full Synthesis system was designed, implemented and technically validated across a range of industrial and field-relevant use cases. This work resulted in a complete solution capable of managing large numbers of machine-learning models on distributed devices without requiring access to their underlying data.

All core components of the system were fully developed. This includes a model-agnostic federated learning method that allows different types of models to be improved collectively, techniques for summarizing data directly on the device, the generation of synthetic datasets in the cloud and advanced model-similarity metrics that enable clustering and anomaly detection. These capabilities make it possible to identify under-performing devices, detect deviations in model behavior and analyze entire fleets based on model characteristics alone. They also ensure that models can be monitored, compared and improved over time, supporting an effective lifecycle-management process across distributed deployments.

A major achievement is the completed real-time dashboard, which supports fleets of more than 10,000 models and provides an action-oriented workflow for users. The interface enables exploration of fleet-level behavior, comparison of models against typical patterns, monitoring of model evolution over time and flexible organization of devices through tag hierarchies. Clustering, outlier detection and seed-model management are fully integrated into a responsive and scalable user experience.

The system was evaluated through multiple pilot studies using real operational data in two main sectors: manufacturing and industrial equipment and energy and thermal systems. These pilots confirmed that the system can detect deviations early, support model-based diagnostics and provide valuable insights at both fleet and device level. The results show that Synthesis is technically mature, scalable and ready for further deployment and commercialization.

Progress beyond the state of the art and expected potential impact (including the socio-economic impact and the wider societal implications of the project so far)

The Ekkono FLIIoT project achieved several results that advance the state of the art in distributed and privacy-preserving machine learning. A key scientific contribution is the development of an advanced knowledge-distillation technique for federated learning that enables heterogeneous models to be improved collectively without sharing raw data. This approach allows different model types to contribute to a shared learning process, helping create stronger starting models and improving performance even when devices operate under diverse conditions.

Another major advance is the development of model-based anomaly detection and clustering techniques that rely on comparing model behaviour rather than accessing underlying device data. Central to this work is a set of similarity metrics that make it possible to identify atypical or drifting models, detect under-performing devices and group equipment with similar operating characteristics, all while keeping data on the device.

Building on these contributions and additional developments in data summarisation, synthetic data generation, model comparison and large-scale orchestration, the project delivered Synthesis, a complete system for managing machine-learning models directly on distributed devices. Synthesis brings these advances together into a scalable solution that can monitor, compare and improve thousands of models across industrial and energy-related equipment without requiring access to raw operational data.

Together, these results demonstrate new possibilities for deploying, analysing and maintaining machine-learning models at scale in contexts where data cannot be centralised. They offer strong potential to support more efficient, reliable and transparent operation across a wide range of real-world deployments.

Synthesis – Clustering dashboard (red cluster showing devices with low error)

Synthesis - Device Funnel Dashboard

This figure illustrates the Synthesis framework

Device inspection – Attribute sensitivity and local error over time (seed model in green)