In modern manufacturing, predictive maintenance has become a cornerstone for ensuring system reliability, reducing downtime, and improving efficiency. Within MODAPTO, one of our PhD researchers, from Université de Lorraine, is exploring a cutting-edge approach: applying Federated Learning (FL) to diagnostics and prognostics in modular and reconfigurable manufacturing systems.
This work was recently presented at the 15th IFAC Workshop on Intelligent Manufacturing Systems, through the paper: Federated Learning for Remaining Useful Life Prediction: A Literature Review
Why does this matter?
Traditional machine learning models require large and centralised datasets to train effectively. In industry, however, data is often fragmented across factories, machines, and companies. Sharing such sensitive data is rarely possible due to confidentiality agreements, competition, and cybersecurity concerns. At the same time, relevant failure data is scarce, as modern machines are more reliable and major breakdowns relatively uncommon.
Federated Learning offers a breakthrough. Instead of moving data, FL enables each company or production site to train models locally on its own data. Only the model parameters—not the raw data—are shared with a central server, which then aggregates them into a global model. In this way, companies benefit from collective intelligence without compromising privacy or business confidentiality.
In predictive maintenance, this means that multiple manufacturers using similar equipment—such as machine tools, motors, or robotics—could collaborate to build robust models that estimate the Remaining Useful Life (RUL) of components. Even if each site has only a few examples of failures, federated approaches make it possible to learn from the distributed experience of the entire network. The result: predictive models that are more accurate, generalisable, and practical for decision-making in real industrial environments.
The MODAPTO PhD project is addressing the main scientific challenges of applying FL to predictive maintenance:
-
Contextual variability: even identical machines behave differently depending on usage, environment, and maintenance history, so models must adapt to this diversity.
-
Few-shot learning: each site often has very limited labelled failure data, requiring strategies that can learn from scarce examples.
-
Convergence and stability: ensuring that the global model remains reliable when trained on heterogeneous sources.
These challenges are not only theoretical. The algorithms may be applied in real industrial pilot sites, including SEW-USOCOME, where predictive maintenance is critical for both efficiency and safety.
To complement this research, the doctoral student has also authored a literature review on Federated Learning for RUL prediction. The study highlights a fast-growing body of work since 2020, experimenting with advanced architectures such as LSTMs, CNNs, Transformers, and Graph Neural Networks. Yet, key challenges remain: current FL solutions struggle with scalability when many clients are involved, often underperform in highly heterogeneous settings, and demand significant computational and communication resources.
By combining this state-of-the-art knowledge with real-world experimentation in MODAPTO, the PhD research is paving the way for privacy-preserving, collaborative, and scalable predictive maintenance solutions. Such advances could help industrial partners anticipate failures earlier, optimise maintenance schedules, and extend the lifetime of critical assets—all while safeguarding sensitive data.
This work shows how MODAPTO goes beyond academic research, delivering tangible impact for the factories of the future by enabling manufacturing systems that are more resilient and sustainable.