Pipe failure modelling for water distribution networks using boosted decision trees

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 ...

Section: Discussionmentioning

confidence: 89%

Sewer Condition Prediction and Analysis of Explanatory Factors

Laakso

Kokkonen

Mellin

et al. 2018

“…Hence, the application of statistical and ML models for pipe failure modeling constitutes an important tool for planning proactive rehabilitation strategies of WDNs. Even in limited data availability, predictive failure models can provide valuable information, helping to prioritize system rehabilitation [9].Predictive models can be classified into physical [10], statistical [11], and data-driven models [6]. In order to predict pipes' propensity to breakage, physical models analyze the load applied to the pipes and their capacity to resist it along with the corrosion that appears on the internal and external walls [10,12].…”

mentioning

confidence: 99%

mentioning

confidence: 99%

Comparison of Statistical and Machine Learning Models for Pipe Failure Modeling in Water Distribution Networks

Giraldo-González

Rodríguez

2020

The application of statistical and Machine Learning models plays a critical role in planning and decision support processes for efficient and reliable Water Distribution Network (WDN) management. Failure models can provide valuable information for prioritizing system rehabilitation even in data scarcity scenarios, such as developing countries. Few studies have analyzed the performance of more than two models, and examples of case studies in developing countries are insufficient. This study compares various statistical and Machine Learning models to provide useful information to practitioners for the selection of a suitable pipe failure model according to information availability and network characteristics. Three statistical models (i.e., Linear, Poisson, and Evolutionary Polynomial Regressions) were used for failure prediction in groups of pipes. Machine Learning approaches, particularly Gradient-Boosted Tree (GBT), Bayes, Support Vector Machines and Artificial Neuronal Networks (ANNs), were compared in predicting individual pipe failure rates. The proposed approach was applied to a WDN in Bogotá (Colombia). The statistical models showed an acceptable performance (R2 between 0.695 and 0.927), but the Poisson Regression was the most suitable for predicting failures in pipes with lower failure rates. Regarding Machine Learning models, Bayes and ANNs exhibited low performance in the prediction of pipe failure condition. The GBT approach had the best performing classifier.

“…The use of these tools made it possible to classify the operating statuses of the water supply systems with an accuracy of 90%. Despite the fact that neural network, naïve Bayes classifier [31] and decision tree [23] methods are used for the detection of working conditions of water supply networks, machine learning methods, based on the empirical factor that can be considered a base for real-time monitoring in the future, have not been used. The research results may significantly contribute to diagnosing operating conditions of water supply networks and may constitute one of the tools supporting the assessment of water supply network conditions in water supply companies.…”

Section: Discussionmentioning

confidence: 99%

“…The lowest accuracy (78.51%) used a k-nearest neighbor algorithm. Neural networks [22] and decision trees [23] were also used to detect operation status of water supply systems.…”

Section: Introductionmentioning

confidence: 99%

The Analysis of Water Supply Operating Conditions Systems by Means of Empirical Exponents

Stańczyk

Burszta‐Adamiak

2019

The stochastic character of water consumption by consumers and the technical condition of water supply systems are the main deterministic random factors influencing the observed changes in flow rate and pressure. The implementation of Supervisory Control and Data Acquisition (SCADA) systems resulted in the creation of dispersed data sets coming from the devices controlling the operation of the water supply system. Thanks to the use of metadata and advanced computer systems of analysis, data from various sources can be analyzed to detect the operating conditions of the water supply system. The aim of the research was to analyze an empirical exponent, determined on the basis of flow rate and pressure measurements for one of the District Metered Areas (DMAs). Modern supervised and unsupervised machine learning systems were implemented to classify the obtained results. The results of the research showed that on the basis of the established empirical exponent in the systems in which the pressure is reduced at night, it is possible to qualify the operating conditions of the water supply system in the DMA with accuracy of up to 90%. The conducted tests may be implemented as a component of expert diagnostic systems in water companies.