Abstract
Modern infrastructures, machines, and manufacturing processes require effective management under constrained resources, making it essential to determine when and how to intervene. Prognostics and Health Management (PHM) offers a systematic framework that leverages data collection and computational models to support the management of nearly any engineering component or system. This dissertation addresses three main PHM areas—Reliability Modelling, Markov Process-based Prognostics, and Maintenance Optimisation—emphasizing data-driven techniques to automate model development and deployment.
Part I focuses on Reliability Modelling, specifically the automated inference of Fault Tree (FT) models from failure datasets. Traditional FT construction often involves iterative collaboration between experts and modellers, which can be prone to error and yield incomplete FT structures. To overcome these limitations, this work proposes the use of Multi-Objective Evolutionary Algorithms (MOEAs) to frame FT inference as a multi-objective task, resulting in the FT-MOEA algorithm. Although FT-MOEA generates compact FT structures, scalability challenges arise. To address this, the SymLearn toolchain uses a divide-and-conquer strategy by detecting modules and symmetries in failure data, while the FT-MOEA-CM extension adds confusion matrix metrics to improve both robustness and scalability.
Part II examines Markov Process-based Prognostics, specifically to model the stochastic deterioration of sewer mains—critical infrastructure that presents monitoring challenges due to scale and limited inspection capacity. Markov chains are employed to represent damage severity levels, evaluated through a Dutch sewer network case study. Both homogeneous and inhomogeneous-time Markov chains are tested, with the Turnbull estimator handling interval-censored data. Results indicate that inhomogeneous-time Markov chains can capture non-linear deterioration patterns more effectively, though care is required to avoid overfitting.
Finally, Part III addresses Maintenance Optimisation of sewer mains, highlighting the complexity of deriving optimal policies due to factors like system scale, data availability, and deterioration model assumptions. Reinforcement Learning (RL) remains relatively underexplored in this context, prompting a Deep Reinforcement Learning (DRL) approach for multi-state deteriorating components. By framing maintenance as a sequential decision-making problem, DRL-based agents achieve better results than heuristics, suggesting that such methods can enhance maintenance policies despite the challenges involved in training these models.
Part I focuses on Reliability Modelling, specifically the automated inference of Fault Tree (FT) models from failure datasets. Traditional FT construction often involves iterative collaboration between experts and modellers, which can be prone to error and yield incomplete FT structures. To overcome these limitations, this work proposes the use of Multi-Objective Evolutionary Algorithms (MOEAs) to frame FT inference as a multi-objective task, resulting in the FT-MOEA algorithm. Although FT-MOEA generates compact FT structures, scalability challenges arise. To address this, the SymLearn toolchain uses a divide-and-conquer strategy by detecting modules and symmetries in failure data, while the FT-MOEA-CM extension adds confusion matrix metrics to improve both robustness and scalability.
Part II examines Markov Process-based Prognostics, specifically to model the stochastic deterioration of sewer mains—critical infrastructure that presents monitoring challenges due to scale and limited inspection capacity. Markov chains are employed to represent damage severity levels, evaluated through a Dutch sewer network case study. Both homogeneous and inhomogeneous-time Markov chains are tested, with the Turnbull estimator handling interval-censored data. Results indicate that inhomogeneous-time Markov chains can capture non-linear deterioration patterns more effectively, though care is required to avoid overfitting.
Finally, Part III addresses Maintenance Optimisation of sewer mains, highlighting the complexity of deriving optimal policies due to factors like system scale, data availability, and deterioration model assumptions. Reinforcement Learning (RL) remains relatively underexplored in this context, prompting a Deep Reinforcement Learning (DRL) approach for multi-state deteriorating components. By framing maintenance as a sequential decision-making problem, DRL-based agents achieve better results than heuristics, suggesting that such methods can enhance maintenance policies despite the challenges involved in training these models.
Original language | English |
---|---|
Qualification | Doctor of Philosophy |
Awarding Institution |
|
Supervisors/Advisors |
|
Award date | 7 Feb 2025 |
Place of Publication | Enschede |
Publisher | |
Print ISBNs | 978-90-365-6406-9 |
Electronic ISBNs | 978-90-365-6407-6 |
DOIs | |
Publication status | Published - Feb 2025 |