Review of sewer deterioration models

Review of sewer deterioration models

The adoption of decision support tools for the definition of cost-effective strategies is seen to gain more importance in the coming years. This development is due for one part to the general degradation of the existing systems and for the other part to changes into the regulations and demands for more transparency in decision-making (Ana and Bauwens, 2007). A key element of decision support systems is the ability to assess and predict the remaining life of the assets (Marlow et al., 2009). For this purpose, deterioration models have been developed to understand and describe the sewer aging based on available CCTV inspections and a list of factors that influence the deterioration. This report first describes the potential sewer deterioration factors and analyzes a panel of literature case studies regarding the relevance of each factor on sewer deterioration. Results are hardly directly comparable, because of the different construction practices, historical backgrounds and environmental conditions of the networks investigated. However, some trends regarding the most significant factors may be identified. In most studies, the construction year and the material seem to be the most relevant factor to explain sewer aging. Pipe size, depth, location and sewer function show generally a medium significance on sewer deterioration. Pipe slope was found to have a low significance for the structural deterioration, but a high relevance on the hydraulic deterioration. The effect of other factors as pipe shape, pipe length, soil type, sewer bedding, presence of trees, installation method, standard of workmanship, joint type, and ground water level have been highlighted but rarely or never investigated. On a second step, this report presents three main approaches for sewer deterioration modeling: deterministic, statistical and artificial intelligence based models. The models can be further categorized into pipe group and pipe level models (Ana and Bauwens, 2010). Pipe group models (e.g. Cohort survival or Markov) can be used to predict the condition of a group of sewers or cohorts and are useful to support strategic asset management, i.e. the definition of long term strategies and budget requirements. These models enable to evaluate the efficiency of several scenarios at the network scale. Pipe level models (e.g. regression, discriminant analysis, neural networks) can be used to simulate the condition of each single pipe. They may be useful to set priorities and justify asset management operations. Pipe level models are tools that can support the utilities in the short and mid-term planning and determine at a finer resolution how, when, and where to rehabilitate sewers. Literature results indicate that cohort survival and Markov models are two useful approaches for modeling the degradation of pipe groups. However, the quality of prediction of these models depends highly on the availability of a large amount of inspection data. Extensive datasets are required to create representative sewer groups (cohorts) with sufficient inspected sewers in each condition state. Regression and Discriminant Analysis were tested on several case studies but showed pretty low prediction performances. Three main reasons could be (i) the non-validity of model assumptions, (ii) the biased distribution of the datasets in terms of number of samples for each condition state and (iii) the lack of data for important deterioration factors. Neural networks have proven to be successful tools for the prediction of the deterioration of individual pipes. However, they require (i) relatively complex and time-consuming training processes and (ii) extensive datasets of CCTV inspection and deterioration factors. Only very few case studies intended to evaluate the quality of prediction of these deterioration models. Furthermore, validation results are often contradictory and hardly comparable since (i) the data available for model calibration differ (percentage of CCTV available, type of deterioration factors available) and (ii) the metrics of the methodologies used to assess the quality of prediction differ. Thus, there is still no clear conclusion about the best modeling approach depending on the modeling purpose (pipe group or pipe level). There is also no clear conclusion regarding the quality of prediction that can be reached since in most case studies only a few percentages of CCTV data were available and many data regarding potential deterioration factors were missing. Further research work is needed in order to (i) identify the most appropriate modeling approach depending on the modeling purpose, (ii) understand the influence of CCTV data availability on the modeling results, (iii) analyze the influence of input data uncertainty (CCTV and deterioration factors) on the modeling processes and (iv) find out the optimum input data requirement (availability of CCTV data and deterioration factors) for model calibration.