Multi-agent reinforcement learning approach for predictive maintenance of a Smart Building lighting system

This paper presents a predictive maintenance methodology for Smart Building systems using fault tree models and Weibull distributions to estimate
component failure probabilities. We introduce connection events to reduce
the complexity of the fault tree architecture. These new events allow us
to capture system interactions and identify critical components. Reinforcement learning-based algorithms are employed to develop maintenance strategies that comply with Quality of Service constraints. A comparative analysis with the Simulated Annealing algorithm reveals that the Reinforcement
Learning algorithms strike a balance between reliability and maintenance
costs, whereas the Simulated Annealing algorithm prioritizes cost efficiency.
The approach is adaptable to any system and requires minimal failure data
for real-world applications, which is one of the primary challenges in Smart
Building maintenance.