1 1.6. Thesis outline 21 evaluating and comparing the assumption of time homogeneity, contrasting homogeneous and inhomogeneous time Markov chains (see Chapter 6). Furthermore, we address interval-censoring (Duchesne, Beardsell, Villeneuve, et al., 2013), a natural condition of sewer inspection data, using non-parametric estimators such as the Turnbull estimator. Phase III focuses on policy optimisation of multi-state components (Part III), building upon the degradation models developed in Phase II (see relations in Figure 1.12). To address this, we formulate the sequential decision problem using Markov Decision Process and employ Deep Reinforcement Learning to create agents capable of advising near-optimal maintenance policies (see Chapter 7). Particular attention is given to evaluating the impact of di!erent degradation model assumptions on maintenance policies and comparing agent performance against well-known heuristics commonly applied in predictive maintenance. 1.6 Thesis outline Figure 1.13 provides an overview of the sections and their respective chapters. Chapter 1 introduces and frames the research. Chapters 2 to 7 consist of peerreviewed published journal or conference papers. Chapter 8 presents the general discussion, and Chapter 9 o!ers the conclusion and recommendations. The primary contributions of this dissertation are outlined in the following section. 1.7 Main contributions Contributions on Reliability Modelling: Data-driven Inference of Fault Tree models In Part I of this dissertation, we explored, for the first time, Multi-Objective Evolutionary Algorithms (MOEAs) to automatically infer FTs from failure datasets. In the domain of reliability modelling, our contributions are three-fold: 1. The FT-MOEA algorithm (Chapter 2), based on an MOEA, accounts for three optimisation metrics, including minimising FT size and accuracy-related error metrics. With FT-MOEA, we can consistently obtain compact FT structures. Data and implementations are available at zenodo.org/record/5536431. 2. The SymLearn toolchain (Chapter 3) employs a “divide and conquer” strategy, exploiting symmetries and modules that may be present in the failure dataset. With SymLearn, we can handle larger problems and thus improve scalability. Data and implementations are available at zenodo.org/record/5571811. 3. The FT-MOEA-CM extension (Chapter 4) expands the multi-objective optimisation function by incorporating metrics computed from the confusion matrix. With FT-MOEA-CM, we improved robustness by consistently achieving global optima for larger problems. Data and implementations are available at https://gitlab.utwente.nl/fmt/fault-trees/ft-moea.
RkJQdWJsaXNoZXIy MjY0ODMw