187 Chapter 9 Conclusions & Recommendations 9.1 Conclusions In Part I of this thesis, we investigated how to obtain e!cient and compact Fault Tree models from failure datasets in a robust and scalable manner and concluded that, with available failure datasets, this task can be e!ectively addressed using FT-MOEA-CM, our proposal based on NSGA-II, a Multi-Objective Evolutionary Algorithm. Our findings indicate that six metrics from the confusion matrix, along with the size of the Fault Tree, are crucial for consistently achieving compact and e"cient Fault Tree structures, ensuring e!ective algorithm convergence. This is crucial for unveiling useful patterns associated with failure propagation through basic and intermediate events in a Fault Tree. The evidence shows that FT-MOEA-CM is more robust and scalable than state-of-theart algorithms such as FT-EA and its predecessor FT-MOEA. Furthermore, in cases involving symmetries in the failure dataset, coupling FT-MOEA-CM with toolchains such as SymLearn may further improve convergence and scalability. In Part II of this dissertation, we investigated how and to what extent it is possible to accurately model Multi-State Deterioration with applications in sewer mains. For this, we explore the use of Markov chains and examined both data availability and model assumptions. High-quality, publicly available inspection datasets for sewer mains are scarce, limiting predictive modelling and validation e!orts. To address this, we shared part of the data from a real-world case study in the Netherlands, contributing to the understanding of sewer mains’ degradation behaviour. We evaluated two typical Markov chain structures to model damage severity levels, concluding that the simpler structure su"ces for typical sewer main degradation. Comparing homogeneous and inhomogeneous time Markov chains, our results suggest that inhomogeneous chains better capture non-linear stochastic behaviours. However, inhomogeneous chains introduced additional hyper-parameters, leading to overfitting and hindering predictability. For calibration, we used a process combining the Metropolis-Hastings algorithm with Sequential Least Squares Programming.
RkJQdWJsaXNoZXIy MjY0ODMw