184 Chapter 8. Discussion The potential of DRL is significant, highlighted by growing interest evidenced in recent reviews (Real Torres, Andreiana, Ojeda Roldán, et al., 2022; Siraskar, Kumar, Patil, et al., 2023; Marugán, 2023; Li, Zheng, Yin, et al., 2023) and the unexplored applications in infrastructure systems such as sewer networks. What are the implications of using DRL for the maintenance of large multi-component systems? DRL creates virtual agents that, when well-trained, act as virtual ‘experts’ exposed to diverse scenarios described by the environment. The knowledge encoded in these agents can potentially improve strategic maintenance. Though challenging, this could eventually become key in managing infrastructure and resource constraints, enabling timely actions and optimal resource allocation. In Chapter 7, we demonstrate how these agents outperform heuristics. An implication of this work is for instance, that the knowledge acquired by the trained agents potentially could improve heuristics, or even at a higher level, understanding the agent’s behaviour could o!er valuable insights for strategic maintenance. Additionally, we observed that agents developed distinct behaviours in line with the dynamics of their environments, emphasising the importance of properly integrating prognostics within the policy optimisation framework. 8.4 Moving towards comprehensive Prognostics and Health Management: Closing Thoughts To recap, this dissertation explored several key facets of the Prognostics and Health Management (PHM) paradigm, deploying advanced methodologies to support reliability analysis such as Multi-Objective Evolutionary Algorithms for automatically constructing Fault Tree models from data, and Markov chains for Multi-State Deterioration Models (MSDMs) applied to sewer mains. We also delved into strategic maintenance planning using Deep Reinforcement Learning (DRL), showcasing the versatility yet complexity of DRL in crafting optimal maintenance strategies. Reflecting on my research journey, I find that e!ective and comprehensive PHM remains a significant challenge. A key issue is the gap between the expectation of what PHM can o!er and what is achievable with current technology. During my PhD, I observed many companies showing interest in predictive maintenance systems but lacking the actionable data needed for e!ective implementation. To close this gap, it is crucial, among others, to align data collection strategies with specific model requirements, especially in prognostics, where abstracting long-term degradation profiles is di"cult. Collecting data without clear objectives is ine"cient and wastes valuable resources.
RkJQdWJsaXNoZXIy MjY0ODMw