Industry 4.0 enabled suppliers and manufacturers to introduce new technologies concepts. The mains concepts introduced are the Internet of Things (IoT), cloud computing and data analysis, to know about machine’s states and detect potential defects. This approach is called “predictive maintenance”, it evaluates real conditions and their evolution to predict eventual maintenance operation before breakdown.
However, predictive maintenance deployment at an industrial scale gives rise to several questions about: sensor data trust, required algorithm to detect defects and predict breakdown, its adaptability to large volume and to equipment.
To be implemented, predictive maintenance use IoT technology to collect useful measures to detect failures. Then, these data can be transfer to the central server to be stocked and analyzed (which involves latency and important energy costs), or they can be stocked and processed in the periphery network which is called “edge computing”.
Sensor diversity highlights interrogation about a large scale implementation. Indeed, approaches operating on local learning do not make model generalization easier as it generates execution diversity in situations, it also makes more difficult “cold start” because models are hardly generalizable due to the variety of use contexts.
In this context, a new learning approach called “federated learning” emerged. This approach enables to several nodes, with various data and resources, to take part to a global model training. Each node at an EDGE level contributes to the generalized learning process by calculating independently a model parameter with the local data.
In this context, Berger-Levrault wants to develop a predictive maintenance approach which exploits the federated learning collaborative approach and take advantage of the edge computing. The goal is to propose a monitoring system to:
- Enable “cold start” when a new equipment is installed,
- Detect anomalies learned is a shared way to all clients.
That’s why, Hamza Safri, young telecommunication engineer has join the DRIT to start his recently accepted CIFRE thesis named: “Federated learning for predictive maintenance”.
He’s doing his thesis in collaboration with Grenoble INRIA’s laboratory and his thesis director and supervisor are Frédéric Desprez and Denis Trystram.
