Our colleague Kévin Ducharlet, a researcher in the BLPredict team, has just defended his PhD thesis in Computer Science and Control on September 30, 2023!
After three years of applied research in partnership with CARL Berger-Levrault, and the Laboratoire d’Analyse et d’Architecture des Systèmes (LAAS-CNRS) in Toulouse, Kévin continues the BLPredict adventure as an expert in unsupervised anomaly detection! His in-depth knowledge of this field enables him to optimize the reliability of measurements taken in sensor networks, for optimized predictive maintenance.
The quest for reliability in industrial equipment control
Its R&D work focuses on the new ways of managing equipment and machinery enabled by the transformations of Industry 4.0. With the development of the Internet of Things (IoT), maintenance deployment now relies on the deployment of sensor networks. These small, battery-powered devices, with their low computing and memory capacities, collect a mass of data, process it and transmit it over a wireless network.
However, the challenge lies in optimizing the reliability of the measurements made within these networks, while ensuring automated deployment of the system at the customer’s premises. And that’s what his thesis is all about. Kévin has contributed to our knowledge of the unsupervised anomaly detection approach. This technique, which consists in learning a concept of normality from past measurements in order to deduce what is abnormal in the future, will certainly enrich the capabilities offered by predictive maintenance.
Kévin has developed several devices aimed at making anomaly detection and resolution more precise and accessible:
Our shared ambition: to optimize vehicle fleet management, vehicle telematics, transport and goods monitoring, and improve supply chain and facility management processes.
- The WOLF (Wireless OutLier detection Framework) operational framework, which structures knowledge of industrial anomalies. This toolbox helps maintenance managers to resolve a whole range of anomalies, in a personalized way according to the management methods used by the manufacturer.
- The WOLF-Eval evaluation approach integrated into the WOLF operational framework, which calculates the accuracy of available methods according to the definition given to the anomaly.
- The DyCF and DyCG anomaly detection methods, which require very little parameterization by engineers, for accessible and rapidly deployable unsupervised maintenance, offer satisfactory results.
Applying these innovations to BLPredict would offer maximum reliability of data from sensor networks, without requiring the intervention of experts on the customer’s side: the assurance of complete, automated management system autonomy! And given the growing number of industrial customers who place their trust in us (and their wide range of sectors!), our 100% industry-specific IIoT platform will be able to extend to a wide range of applications.
Would you like a quick overview of unsupervised anomaly detection methods applied to data flows? Read the full article here: https: //hal.science/hal-03765550v1
Kévin Ducharlet, Louise Travé-Massuyès, Marie-Véronique Le Lann, Youssef Miloudi. Study of unsupervised anomaly detection methods applied to data streams. 20th Rencontres des Jeunes Chercheurs en Intelligence Artificielle (RJCIA 2022), Jun 2022, Saint-Etienne, France. ⟨hal-03765550⟩
