Maintenance models to provide better performance of steelmaking production lines that make use of recycled scrap

Abstract

To meet growing market demands and remain competitive, modern production systems are widely adopting technological innovations, such as systems monitoring and machine connectivity, which leads a huge amount of data available about the health of the system. In this scenario, condition-based maintenance can be a powerful tool for industry competitiveness due to its ability to intervene in the system in real-time by its condition monitoring, enhancing the system availability, reliability, and cost when compared with time-based maintenance policy. However, the large amount, variety, and dimensionality of the data that comes from a production line create a problem with a large space of states, which is intractable with traditional maintenance models. To overcome this challenge, emerging tools and methodologies of the areas of Artificial Intelligence and Machine Learning are being used in the maintenance planning. Which Deep Reinforcement Learning (DRL) proved to be efficient for maintenance decision making based on multiple component conditions of a production line. Therefore, this work proposes two maintenance models: an opportunistic maintenance model considering production data to anticipate maintenance actions, and a DRL-based model to support the decision-maker in making optimal maintenance decisions in a serial production line based on system monitoring. The environment under study was a steelmaking production line. A simulation model was built to represent and simulate the behavior of the system. In the DRL model, two scenarios regarding distinct aspects of the system were investigated. A DRL framework was constructed for each scenario to learn through interaction between an agent and the simulated environment the optimal maintenance policy. Both models use as a decision criterion the minimization of the expected long-run cost rate. To evaluate the proposed models, a numerical case study was performed. The sensitivity analysis of the models was also performed to observe their behavior in the face of variations in the system parameters. As result, the models behave as expected and the proposed policies show a better result in terms of cost, system availability, and production in comparison with other time-based policies used in the steel context.