Born and raised in Grenoble, Guillaume studied at the Ecole des Ingénieurs du Numérique (ISEN) in Lille, where he specialized in computer science and electronics. After a professional contract in artificial intelligence, he chose to return to Grenoble to pursue a thesis at CEA-Leti on predictive maintenance using ultrasound.
“The subject was exactly what I was looking for," he explains. “It was a real passion project and an opportunity to discover one of the largest laboratories in the region."
At the PHM international conference, Guillaume presented his paper entitled “Knowledge-informed symbolic regression for new features discovery for degradation analysis of rolling bearings," which earned him the best paper award. His work is based on analyzing ball bearing vibrations to assess their degradation, exploring an innovative approach: applying symbolic regression to predictive maintenance.
Symbolic regression: combining physics and artificial intelligence
Ball bearings are ubiquitous in industry, and are both essential and fragile: 10 billion are produced each year, with nearly 50 million failures recorded. These failures can be caused by wear, lubrication problems, or contamination. It is in this context that Guillaume is focusing on developing a monitoring method to anticipate such damage. In order to do this, he records the vibrations of the bearings at regular intervals, which he then analyzes to detect any abnormal changes.
His approach is based on symbolic regression, an artificial intelligence technique that automatically generates the mathematical equations best suited to the observed data. Guillaume explains that “based on certain assumptions, it is possible to form an analytical model and monitor the behavior of deterioration," which makes it possible to discover characteristics representative of bearing aging.
He has also demonstrated that combining this method with traditional deep learning approaches yields more effective results, while remaining lightweight enough to be embedded in systems.
Symbolic regression at the service of the industrial sector
Guillaume's goal is clear: developing a predictive model compact enough to be embedded directly in microcontrollers. This would enable real-time monitoring of a machine, while reducing the need for hardware resources.
Such advances could benefit a wide range of industrial applications: electric motors, pumps, wind turbines, and more broadly any system that uses ball bearings and is susceptible to vibration due to wear and tear. Ultimately, these tools would extend the lifespan of machines, prevent unplanned downtime, and optimize maintenance in key sectors.
Building on this research, Guillaume also won the award for best poster at IN4PL 2025, an international conference on innovation in industrial production and logistics. His study, “Feature Fusion with Online Principal Component Analysis for Embedded Unsupervised Machine Monitoring," presents a lightweight, unsupervised method for continuously monitoring the condition of bearings through vibration analysis.