7 7.8. Discussion and conclusions 171 scenario is plausible, as new pipes can exhibit high severity levels at a young age due to defects in the material or errors during the construction and installation process. This concept is represented in the MSDM by the initial probability state vector (p (0) k ). Additionally, Agent-G recommends maintenance at the interval when the pipe reaches the age of 26 years; (b) Agent-E suggests replacement at approximately 62 years, without recommending further maintenance; (c) CBM advocates for maintenance at about 65 years, followed by replacement at 70 years, in line with heuristics described in Section 7.6.2; (d) SchM consistently performs maintenance at regular intervals, yet faces significant deterioration, culminating in failure around 97 years. 7.8 Discussion and conclusions In this chapter, we explore the applications of Prognostics and Health Management in sewer main asset management. Our study focuses on component-level (i.e., pipe-level) maintenance policy optimisation by integrating stochastic multi-state deterioration modelling and Deep Reinforcement Learning (DRL). The goal is to assess the e!ectiveness of DRL in deriving cost-e!ective maintenance strategies tailored to the specific conditions and requirements of sewer mains. A key contribution of our work is the integration of prognostics models with a maintenance policy optimisation framework. Our reward function aligns with damage severity levels, enabling a more complex and realistic maintenance optimisation setup. Our methodology includes a real-world case study from a Dutch sewer network, which provides historical inspection data. Through hyper-parameter tuning and policy analysis, we benchmark our optimised policies against traditional heuristics, including condition-based, scheduled, and reactive maintenance. Our findings suggest that agents trained with the Proximal Policy Optimisation algorithm are highly capable of developing strategic maintenance policies, adapting to pipe age, and surpassing heuristic baselines by learning cost-e!ective dynamic management strategies. To evaluate the impact of deterioration model assumptions, we trained one agent using the Gompertz probability density function and another using the Exponential probability density function. During testing, both agents were assessed in an environment parametrised with the Weibull probability density function. The Gompertz-trained agent, whose behaviour more closely resembled the Weibull model, demonstrated better generalisation, resulting in more e!ective maintenance policies compared to the Exponentialtrained agent. Future work: The following directions are identified:
RkJQdWJsaXNoZXIy MjY0ODMw