Deep reinforcement learning-based maintenance decision-making for a steel production line - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2021

Deep reinforcement learning-based maintenance decision-making for a steel production line

Résumé

In the 4.0 industry, the adoption of system monitoring technologies provides a large amount of data about the health of the system, which raises a challenge to adopt condition-based maintenance (CBM). Due to its capability to act into the system in real-time based on its embedded condition monitoring equipment, which can help to reduce O&M cost and enhancing the system availability, CBM has become a relevant approach for industry competitiveness. However, to take the advantages of huge data in maintenance decision-making, an important issue to be considered is the large space of states and actions, which is difficult, even impossible, to cope with the traditional maintenance model. To overcome this issue, integrating emerging tools of Machine Learning and Artificial Intelligence into maintenance decision-making and optimization seems to be promising. Therefore, this work proposes a Deep Reinforcement Learning (DRL)-based maintenance optimization for a steel production line, in which the maintenance decisions are made based on the real-time data about the system condition. The production line under study uses metal scrap as the raw material for the steelmaking. Before its usage, the scrap needs to be crushed in a shredder machine, which is the most crucial process. An intermediated buffer is used to keep supplying crushed scrap for the remaining stations when the machine is turned off for maintenance actions. A simulation model is built to simulate the dynamic of the production line. A DRL framework is then built to learn through the interactions with the environment in finding the optimal maintenance policy with lowest maintenance cost. A numerical case study is performed to evaluate the proposed DRL maintenance approach comparing with conventional maintenance policies. As result, the proposed DRL approach shows a better result in terms of cost along with the increase of the system availability.
Fichier non déposé

Dates et versions

hal-03379905 , version 1 (15-10-2021)

Identifiants

Citer

Waldomiro Ferreira, Cristiano Cavalcante, Phuc Do Van. Deep reinforcement learning-based maintenance decision-making for a steel production line. 31st European Safety and Reliability Conference, ESREL 2021, Sep 2021, Angers, France. ⟨10.3850/978-981-18-2016-8_600-cd⟩. ⟨hal-03379905⟩
61 Consultations
0 Téléchargements

Altmetric

Partager

Gmail Facebook X LinkedIn More