Sewer condition is commonly assessed using closed-circuit television (CCTV) inspections. In this paper, we combine inspection results, pipe attributes, network data, and data on pipe environment to predict pipe condition and to discover which factors affect it. We apply the random forest algorithm to model pipe condition and assess the variable importance using the Boruta algorithm. We analyse the impact of predictor variables on poor condition using partial dependence plots, which are a valuable technique for this purpose. The results can be used in screening pipes for future inspections and provide insight into the dynamics between predictor variables and poor condition.2 of 17 different pipe attributes and environmental variables affect pipe condition provides an understanding of the mechanisms that cause poor condition. This provides valuable information not only for inspection decisions, but also for future installations. This study focuses on modelling the prevailing condition of a pipe network, and hence lifespan models are discussed only when they address factors explaining pipe deterioration.The methods previously applied for condition modelling include both traditional statistical methods and machine learning methods. Salman and Salem [8] compared the performance of binary logistic regression, multinomial logistic regression, and ordinal regression in modelling sewer condition. They found ordinal regression to be unsuitable for the study, while binary logistic regression provided the best results. Ariaratnam et al. [9], Ana et al. [10], and Fuchs-Hanusch et al. [11] also applied logistic regression to model poor conditions. Chughtai and Zayed [12] applied multiple regression to model a five-level sewer condition scale. They created four different models: three for predicting the structural condition of different materials and one for predicting the operational condition. Savic et al. [13] and Savic et al. [14] applied a different form of regression, a data-driven modelling algorithm called evolutionary polynomial regression, to model the blockage and collapse events in different pipe classes. Khan et al. [7], Tran et al. [15], and Sousa et al. [16] applied neural networks to model how different variables affect the condition of the sewer. Tran et al. [15] modelled the structural condition of stormwater pipes, while Khan et al. [7] modelled sewer condition and Sousa et al. [16] the condition of sanitary sewers. Sousa et al. [16] compared the performance of artificial neural networks (ANNs) and support vector machines (SVMs) with that of logistic regression. They found that ANNs provided the highest classification performance. SVMs, on the other hand, provided excellent results in the study by Mashford et al. [17], who predicted sewer condition using a five-level scale. Decision trees were applied by Syachrani et al. [18] and Harvey and McBean [19], and random forests by Harvey and McBean [20]. Syachrani et al. [18] found that decision trees consistently outperformed the regression and neural networks in ...

Survival models can support the estimation of the resources needed for future renovations of sewer systems. They are particularly useful, when a large share of network will need renovation. This paper studies modelling sewer deterioration in a context, where data are available for pipes selected for inspections due to suspected or experienced poor condition. We compare the random survival forest and the Weibull regression for modelling survival and find that both methods yield similar results, but the random survival forest performs slightly better. We propose a method for estimating the range in which the actual network survival curve lies. We conclude that in order to reach reliable results, a life span model needs to be constructed based on a random sample of pipes, which are then consecutively inspected and in addition, censoring and left truncation need to be accounted for. The inspection data applied in this paper had been collected with the aim of finding pipes in poor condition in the network. As a result, the data were biased towards poor condition and unrepresentative in terms of pipe ages.

No abstract