Adaptative Reinforcement Learning Approach for Predictive Maintenance of a Smart Building Lighting System

Due to advancements in sensing technologies, enhanced IoT architectures, and expanded connectivity options, predictive maintenance has emerged as a compelling solution within the context of Industry 4.0 for industrial systems. However, within this landscape, such as in Smart Buildings (SBs), the lack of failure data poses a significant challenge for implementing traditional data-based approaches documented in the literature. Additionally, SBs are complex systems of systems, where failures in one subsystem can propagate and impact other interconnected systems, adding layers of complexity to maintenance decision-making processes. Lastly, predictive maintenance has tradition-ally been used primarily in the context of operational safety, with the primary goal of avoiding undesirable events at all costs. Only recent work focuses on predictive maintenance to optimize operating costs. In light of these challenges, this paper proposes a Reinforcement Learning approach for Predictive Maintenance tailored for the lighting system within a Smart Building. It aims to minimize costs subject to a Quality of Service constraint by leveraging a Markov Decision Process and Q-learning adaptation, which adjusts based on the life cycle of system components. The results demonstrate that, even with limited system data, meeting defined quality of service standards and minimizing maintenance costs is feasible.