Résumé: The manufacturing world is subject to ever-increasing cost optimization pressures. Maintenance adds to cost and disrupts production; optimized maintenance is therefore of utmost interest. As an autonomous learning mechanism reinforcement learning (RL) is increasingly used to solve complex tasks. While designing an optimal, model-free RL solution for predictive maintenance (PdM) is an attractive proposition, there are several key steps and design elements to be considered-from modeling degradation of the physical equipment to creating RL formulations. In this article, we survey how researchers have applied RL to optimally predict maintenance in diverse forms-from early diagnosis to computing a "health index" to directly suggesting a maintenance action. Contributions of this article include developing a taxonomy for PdM techniques in general and one specifically for RL applied to PdM. We discovered and studied unique techniques and applications by applying (a text mining technique). Furthermore, we systematically studied how researchers have mathematically formulated RL concepts and included some detailed case-studies that help demonstrate the complete flow of applying RL to PdM. Finally, in Sect. 14, we summarize the insights for researchers, and for the industrial practitioner we lay out a simple approach for implementing RL for PdM.

Langue: Anglais