“Predictive maintenance provides a true map of a system's lifecycle by making it possible to identify failures and performance degradation and allowing intervention at the right time and at lower cost," explained Youssof Fassi, a PhD student in signal processing and AI.
“Our strategies are most effective when applied at the system level, where complex, interacting phenomena can be observed, and we address these challenges using artificial intelligence tools, with a very pragmatic approach."
It is no surprise that these enhanced tools are being incorporated in multiple industrial sectors and are receiving recognition.
At the 2025 International Conference on Prognostics and Health Management in Seattle, Guillaume Prevost, a PhD student in signal processing and AI, presented a paper titled, “Knowledge-Informed Symbolic Regression for New Features Discovery for Degradation Analysis of Rolling Bearings." It won a Best Paper Award.
A New AI Paradigm
“The main challenge in predictive maintenance is the lack of historical data and its heterogeneity," Guillaume explained. “Instrumenting a machine is time-consuming and costly, so we developed an approach that combines the processing of machine sensor data with domain expertise and physical models of the system. This is a new paradigm of artificial intelligence, known as physics-informed AI."
Like most CEA-Leti projects, predictive maintenance involves multidisciplinary teams, for example expertise in physical modeling, which Youssof brings, and Guillaume's signal-and-data processing.
A Current Case
Team member Leila Merzak, a PhD student in modeling and signal processing, said one of the team's current use cases is focused on developing digital twins for damage prediction and state of health estimation on mechanical structures. For example, in the framework of her PhD research, on knee prostheses.
“We combine physical models with AI models to predict alignment defects in knee prostheses. This allows us to anticipate mechanical loosening, which can lead to surgical re-intervention procedures that are both complex and costly," she said.
Down the Road: Prescriptive Maintenance
Célestin Ott, a research engineer-multiphysics modeling at CEA-Leti, explained that integrating digital twins with physics-informed artificial intelligence enables more accurate and targeted predictive maintenance by improving the reliability of fault detection and degradation forecasting.
“That is typically true for multi-scale systems. For instance, at the component level within a power converter. It is also much more robust," he said. “The next step is to move toward prescriptive maintenance, which involves providing feedback to the system based on predictions, to optimize its service life. The goal is to move toward an industry that is safer, more efficient, and more sustainable."
'Waiting for Croissants'
Like all well-matched research teams, the members recognize an opportunity to share a humorous moment along with their “very constructive and engaging exchanges", as Célestin describes them.
As when Guillaume ordered a milling machine for the lab to conduct experimental studies on tools' state-of-health (SoH) monitoring in rotating machinery, using ultrasonic sensing to anticipate degradation for predictive-maintenance purposes.
Guillaume: “It arrived as a kit, so I was a bit overwhelmed by it all."
Youseff: “Yes, and on top of that, there were bets going around on how long it would take you to assemble your milling machine."
Liela: “So, we’re still waiting for your croissants.”