PROTEUS · Project

Real-Time Machine Learning for Predicting Defects in Steel and Heavy Industry Production Lines

manufacturingPilotedTRL 7

Imagine a steel factory producing thousands of tonnes of metal every day — and having no way to spot quality problems until it's too late. PROTEUS built a smart early-warning system that watches production data as it flows in, spots patterns that predict failures, and shows operators what's happening in real time on a visual dashboard. Think of it like a weather radar for your factory floor — instead of forecasting rain, it forecasts defects. The technology was tested on a real ArcelorMittal hot strip mill and plugged into Apache Flink, Europe's own big data engine.

By the numbers

EUR 3,156,525

EU funding for development and industrial validation

consortium partners across 3 countries

67%

industry partners in the consortium

SMEs in the consortium

total project deliverables produced

prototype iterations before final demonstrator

The business problem

What needed solving

Heavy industry plants generate enormous volumes of sensor data every second, but most of it is analyzed after the fact — if at all. By the time a quality defect or equipment failure is detected, the damage is already done: scrapped product, unplanned downtime, and wasted energy. Manufacturers need a system that learns from live data streams and predicts problems before they happen, not hours or days later.

The solution

What was built

PROTEUS built SOLMA — a library of scalable online machine learning algorithms integrated into Apache Flink — plus a declarative language for building predictive models on streaming data, automatic drift and anomaly detection, real-time interactive visualization tools, and a final demonstrator validated on ArcelorMittal's operational Hot Strip Mill. The project delivered 34 total deliverables across 3 prototype iterations.

Audience

Who needs this

Steel mills and metals processors with continuous production linesProcess manufacturers (chemicals, paper, glass) generating high-velocity sensor dataIndustrial IoT platform providers looking for proven streaming ML algorithmsEnergy utilities managing real-time grid or plant monitoring dataPredictive maintenance service companies serving heavy industry

Business applications

Who can put this to work

Steel and metals manufacturing

enterprise

Target: Steel mills and metals processors running continuous production lines

If you are a steel manufacturer dealing with unpredictable quality defects on your hot strip mill or continuous casting line — this project developed a real-time machine learning system validated at ArcelorMittal that detects anomalies and predicts failures as data streams in. The system was built on Apache Flink and tested across 3 prototype iterations with a final demonstrator on an operational hot strip mill.

Process manufacturing (chemicals, paper, glass)

enterprise

Target: Process plants with high-speed production lines generating massive sensor data

If you are a process manufacturer struggling to analyze the flood of sensor data from your production lines fast enough to act on it — PROTEUS built a library of scalable online machine learning algorithms that process data streams in real time, including automatic drift and anomaly detection. The system was designed for extremely large data sets and validated in an industrial setting with 6 partners across 3 countries.

Industrial IoT and predictive maintenance

any

Target: Companies offering predictive maintenance or industrial analytics platforms

If you are an industrial analytics provider looking for proven streaming ML algorithms to embed in your platform — PROTEUS created SOLMA, an open library of scalable online machine learning algorithms integrated into Apache Flink, plus a declarative language that makes it easier to build predictive models on live data streams. The library went through 3 versions covering basic streaming algorithms, online learning, and drift detection.

Frequently asked

Quick answers

What would it cost to implement this kind of real-time analytics system?

The PROTEUS project received EUR 3,156,525 in EU funding to develop and validate the technology with 6 partners over 3 years. A commercial deployment would likely cost significantly less since the core algorithms and Apache Flink integration already exist. Licensing or partnership terms would need to be discussed with the consortium.

Has this been tested at industrial scale, not just in a lab?

Yes. The final demonstrator was validated on an operational Hot Strip Mill at ArcelorMittal, one of the world's largest steelmakers. The system went through 3 prototype iterations (V1, V2, V3) before the final industrial validation. This is real factory-floor testing, not a lab simulation.

Who owns the IP and how can I license it?

The technology was developed by a consortium of 6 partners including 3 SMEs and Bournemouth University as coordinator. IP ownership is shared among consortium members. Parts of the work are integrated into Apache Flink, which is open-source, but proprietary components would require licensing discussions with the relevant partners.

Can this work for industries other than steel?

The project objective explicitly states the techniques are 'general, flexible and portable to all data stream-based domains.' While validated on steelmaking, the SOLMA library and Flink extensions are designed for any scenario with high-velocity data streams — energy, chemicals, transport, or telecommunications.

How does this compare to existing predictive analytics tools?

PROTEUS specifically addresses online machine learning — learning from data as it streams in, rather than training on historical batches. It also includes automatic drift detection, meaning the system adapts when production conditions change. These capabilities were integrated directly into Apache Flink, Europe's big data platform.

What's the timeline to get this running in my plant?

The core technology is developed and validated. Integration would depend on your existing data infrastructure and whether you use Apache Flink. Based on available project data, the consortium spent roughly 3 years building and validating 34 deliverables, but deployment of the finished system would be substantially faster.

Consortium

Who built it

The PROTEUS consortium is strongly industry-oriented: 4 out of 6 partners (67%) come from industry, including 3 SMEs, with Bournemouth University coordinating and one research organization rounding out the team. The consortium spans 3 countries (Germany, Spain, UK). The heavy industry presence — likely including ArcelorMittal or its technology suppliers — means the technology was built to meet real production requirements, not just academic benchmarks. For a business buyer, this is a positive signal: the people who built this system had factory floors in mind from day one, and the final demonstrator was tested in an actual steelmaking operation.

BOURNEMOUTH UNIVERSITYCoordinator · UK
TREELOGIC TELEMATICA Y LOGICA RACIONAL PARA LA EMPRESA EUROPEA SLparticipant · ES
DEUTSCHES FORSCHUNGSZENTRUM FUR KUNSTLICHE INTELLIGENZ GMBHparticipant · DE
TRILATERAL RESEARCH LTDparticipant · UK
ARCELORMITTAL INNOVACION INVESTIGACION E INVERSION SLparticipant · ES

How to reach the team

Bournemouth University (UK) coordinated this project. SciTransfer can facilitate an introduction to the research team.

Order a report on this consortium See BOURNEMOUTH UNIVERSITY profile →

Next steps

Talk to the team behind this work.

Want to explore how PROTEUS streaming analytics could work in your production environment? SciTransfer can arrange a direct introduction to the development team and help assess fit for your use case.

Order a report on this topic More in Manufacturing & Industry 4.0