188 Chapter 9. Conclusions & Recommendations Although we did not account for interval censoring during calibration, our results aligned well with the non-parametric Turnbull estimator (which accounts for interval-censoring), suggesting that the e!ects of interval censoring are negligible for this case study. Answering this research question more comprehensively, Markov chains o!er a suitable approach for modelling damage severity levels by representing them as discrete states within the chain. However, the applicability of these models, particularly as implemented in this dissertation, is entirely data-driven, meaning their usability and e!ectiveness heavily rely on data quality. In the case of sewer mains, as previously noted, data quality remains a significant concern. While Markov chain models provide a solid initial step, more advanced degradation modelling techniques are necessary to overcome the limitations imposed by data quality. Finally, in Part III of this dissertation, we investigated how to devise near-optimal maintenance strategies for components with Multi-State Deterioration such as sewer mains using Deep Reinforcement Learning. We evaluated the e!ectiveness of Deep Reinforcement Learning (DRL) in developing cost-e!ective maintenance strategies for sewer mains. A key contribution is the integration of Multi-State Deterioration models within a DRL optimisation framework. We proposed a novel fully-observable Markov Decision Process (MDP) with a reward function aligned to damage severity levels. We benchmarked agent-based policies against traditional heuristics, including condition-based, scheduled, and reactive maintenance. Our results suggest that agents trained with the Proximal Policy Optimisation (PPO) algorithm excel in developing dynamic, cost-e!ective strategies, surpassing heuristic baselines. Our experiments also compared the impact of homogeneous and inhomogeneous assumptions in the deterioration models on the agent’s behaviour. For this, one agent was trained with the Gompertz (inhomogeneous) function and another with the Exponential (homogeneous) function. In testing, the Gompertz-trained agent outperformed the Exponential-trained agent, likely due to the Weibull-distributed deterioration model in the test environment, which aligns more closely with the Gompertz function. 9.2 Recommendations 9.2.1 Automatic Inference of Fault Tree Models While the algorithm and extensions proposed in Chapters 2, 3, and 4 improve FT model inference, several challenges remain. 1. Local optima: Minimising FT model size can result in local optima, as irrelevant structures may satisfy Pareto criteria. For instance, an FT with
RkJQdWJsaXNoZXIy MjY0ODMw