Machine Learning - Based Sorting Nuclear Fuel Waste

INMM & ESARDAAugust 21-26, 2021 – Joint annual meetingMachine Learning - based sorting nuclear fuel waste Stéphane Puydarrieux - Orano Recyclage - La HagueJean-Jacques Dupont – Orano ProjetsHéléna Henry – SiléaneMickael Picq – PICQ ConsultingAbstract:In order to store securely old nuclear wastes from UNGG reactors out of the Silo 130 at Orano La Hague (France), precise sorting means were investigated with the goal of regulating magnesium level in each waste container. Such an endeavour is meant to prevent high level releases of Hydrogen, to avoid the consequences of a too high Hydrogen release during ulterior storage.To sort out the waste from silo and quantify the magnesium, Siléane and Orano have come up with a robotized solution provided with Vision technology (3D Vision system + Laser) and AI (Machine Learning algorithms).After detection, extraction and 3D reconstitution, each part of waste is analyzed by a predetermined tree-based classifier (5 years development and runs using nuclear waste simulants). These algorithms allow to identify the kind of waste that is taken from the silo, whether it is Aluminum, Magnesium or Graphite. After a long period of development, runs and inactive tests to validate the solution, this project has entered a new phase in summer 2019 : after obtaining the ASN authorization mid-2019, the industrial plant started with active waste at the end of 2019 and has produced 5 containers in a fully supervised modus to test the algorithms’ behaviour on the real wastes. Now, a thorough review of the ML-based classifier is currently underway to optimize and qualify the process as active. From April 2021, the machine is due to work with the upgraded AI-based Classifiers in fully autonomous modus. Considered as the latest AI case study in the nuclear industry, applications of AI-based recognition technology to material sorting and waste conditioning looks promising.Keywords:Nuclear waste management, machine learning, 3D images