18 Chapter 1. Introduction larger datasets (e.g., more basic events) without significant resource consumption increases. Robustness is the ability to perform consistently across diverse inputs. A robust algorithm should reliably produce correct and consistent FT models, even with noisy data or when re-evaluated on the same dataset. Completeness is the challenge of generating FT structures that include all failure modes present in the data. Part II: Multi-state deterioration modelling Approaches to model deterioration in sewer main systems can be categorised into three classes: physics-based, Machine Learning (ML), and probabilistic models (Hawari, Alkadour, Elmasry, et al., 2020; Saddiqi, Zhao, Cotterill, et al., 2023). In Part II.3 (page 94), we provide an overview of the related literature. Conceptually, the most robust approach is physics-based models, as they utilise fundamental physics laws to describe deterioration processes. However, due to epistemic uncertainty, applying these models in diverse contexts is challenging, as various factors influencing deterioration may not be considered (Ana and Bauwens, 2010). With the increasing availability of diverse datasets, ML models have potential applications in assessing sewer mains conditions by identifying useful relationships within the data (Zeng, Z. Wang, H. Wang, et al., 2023). However, challenges in the quality of data collected from sewer main systems (Auger, Besnier, Bijnen, et al., 2024) can negatively impact the performance of ML models. Moreover, these models are inadequate for long-term condition assessment (Kantidakis, Putter, Litière, et al., 2023) when they fail to account for properties such as monotonicity—the consistent change (increase or decrease) of deterioration metrics over time due to factors like wear. Even though probabilistic models can also be learned from data, their mathematical foundations make them more suitable for long-term risk assessment (El Morer, Wittek, and Rausch, 2024). Most probabilistic approaches to model deterioration in sewer mains are based on Markov chains (Ana and Bauwens, 2010), where these models explicitly represent damage severity levels through well-defined discrete states, information collected through CCTV cameras and classified based on norms such as the EN 13508:1. As for the use of Markov chains to model the stochastic deterioration of sewer mains, we identified gaps in case studies, Markov chain structure, assumptions, and calibration. For case studies, the research community needs to share existing studies to enhance evidence on sewer main deterioration models (Tscheikner-Gratl, Caradot, Cherqui, et al., 2019). Regarding Markov chains, assumptions about their structure when modelling transitions between severity levels and the time homogeneity assumption—indicating that stochastic transitions between states are time-independent—lack comparisons. Calibrating Markov chains for e"cient convergence and addressing dataset issues, such as data censorship related to the
RkJQdWJsaXNoZXIy MjY0ODMw