Research on pipe burst in water distribution systems: knowledge structure and emerging trends

Water distribution systems (WDSs) are municipal infrastructures to supply treated water for the functioning of societies. However, operational and environmental stresses continuously deteriorate water pipes and induce pipe bursts, which lead to large water losses, pollution risks, and public discontent (Rajani & Kleiner 2001; Folkman 2012). Consequently, it is significant to conduct research on pipe bursts in order to take effective measures for burst control in WDSs.

The basic mechanism of pipe bursts has revealed that stresses caused by internal/external loads are larger than pipe residual stresses. The process from pipe burying to pipe burst is often long in terms of time, and complex involving various factors, e.g., manufacturing quality, installation quality, pipe deterioration, temperature, soil condition, etc. Plenty of research articles are published on the process and control of pipe bursts in WDSs, and several review articles summarized the research progress on the mechanism, detection, and prediction of a pipe burst. With regard to the pipe burst mechanism, Rajani & Kleiner (2001) reviewed physical models used to calculate the frost load, total in-plane stress on pipes, pipe residual stress, and remaining wall thickness. Bergant et al. (2006) reviewed the research progress of water hammers on phenomenon discovery, danger recognition, numerical method development, and models used in commercial software packages. With regard to pipe burst detection, Li et al. (2015) divided burst detection methods into hardware-based and software-based methods and summarized their advantages and disadvantages. Wu & Liu (2017) classified data-driven approaches for pipe burst detection into classification, prediction-classification, and statistical approaches with performance discussion. With regard to pipe burst prediction, Kleiner & Rajani (2001) classified pipe statistical failure prediction models into deterministic and probabilistic models, and summarized governing equations, comparisons, and required data for implementing these statistical models. Nishiyama & Filion (2013) classified pipe failure prediction models into deterministic, probabilistic, and soft computing methods. Since differences exist in the development, adoption, and application of pipe failure prediction models, Scheidegger et al. (2015) presented a framework consisting of model assumptions, detailed data assumptions, and distinct types of probabilistic predictions to guide the selection of pipe failure prediction models. The burst consequences of large-diameter pipes are usually more severe and costly. Wilson et al. (2017) reviewed both statistical and mechanical failure prediction models for pipes with diameters larger than 500 mm. Dawood et al. (2020) reviewed statistical and machine learning models to predict pipe failure published from 2009 − 2019. Barton et al. (2019) summarized the factors affecting pipe failure, and provided advice for factor selection to build pipe failure prediction models.

The above reviews provided descriptions, comparisons, and critiques of the approaches to explore pipe burst mechanism, detection, and prediction. However, a comprehensive review of pipe burst research is lacking and the emerging trends in recent years are not summarized. Therefore, we aim to answer the following two questions in this study. (1) What is the knowledge structure of pipe burst research from 1991 to 2022? (2) How does pipe burst research evolve over time? This study can grasp a full view of current achievement pipe burst research and provide guidance for future research directions and technological development.

Bibliometrics quantitatively analyse patterns in the scientific literature to understand the knowledge structure and emerging trends of a research field. CiteSpace^®, which can produce visual interfaces in special structures as a bibliometrics software, has been widely used in review articles in various research fields. An effective function of CiteSpace^® is reference co-citation analysis. Its basic assumption is that the references often cited together have a stronger bondage than others. They are picked up to form a cluster in which knowledge structure is described and emerging trends can be detected to some extent (Chen et al. 2010, 2012).

We input articles on pipe bursts into CiteSpace^® (6.1 R3) for reference co-citation analysis. Articles published in the period of January1991 − June 2022 in the Web of Science core collection database with the term ‘pipe burst’ in titles, abstracts, or indexing terms were collected. Articles relevant to pipe burst, with the keywords ‘pipe failure’ and ‘pipe break’, were also included. Then, 845 original research and review articles with a total of 16,813 references were acquired and used in the subsequent analysis.

This study is unique in three aspects. First, the full view of the entire pipe burst research field, which has rarely been discussed, is explored. Second, the dataset to be analysed is constructed using 16,813 references of citation index-based expansion, which is always capable of tracking a more comprehensive development history than only 845 articles by keyword search. Third, the knowledge structure and emerging trends are identified by computer software in an objective manner, and professional opinions in a subjective manner.

Detailed information on the 13 clusters of the pipe burst research is shown in Table 1. Size is the number of co-cited references in each cluster. The clusters with larger sizes usually indicate more articles published on these topics and are usually the core topics in the research field. Silhouette reflects the quality of the cluster homogeneity or consistency. All silhouettes of the 13 clusters are larger than 0.7, which is the standard value of high reliability. The mean year is the median publication year of all references in a cluster.

The stresses acting on a pipe in the longitudinal direction are caused by thermal contraction, natural environment, operation, or human impact, while stresses in the transverse direction are caused by operation pressure, water hammer, soil cover load, traffic load, or expansion of frozen moisture (Rajani & Kleiner 2001). If the stresses exceed the limit of pipe residual ultimate strength, a burst would occur. Five clusters (i.e., research topics) are obtained in these research categories as follows.

A water hammer is formed by a hydraulic transient shock that was induced by the start-up or shutdown of pumps or valves. Considerable achievements have been reported on the concepts, causes, processes, and hazards of water hammer by deducing analytical solutions, developing numerical models, establishing physical models, and formulating field tests (Bergant et al. 2006; Riedelmeier et al. 2014; Yao et al. 2016).

The main hazard of water hammers is pipe bursts due to high pressure. Other hazards, including joint loosening due to impact force and pipe wall erosion due to cavitation, deteriorate pipe physical condition and also increase pipe burst risk. Wang et al. (2014) presented a pipe burst risk assessment model based on a water hammer and identified risk factors such as maximum water pressure, maximum vacuum, maximum vapour volume, and maximum transient force. The relationship between water hammer and pipe burst is complex. Then, an in-depth analysis of the role played by the water hammer in a pipe burst is still required.

Pipe structural reliability analysis estimates the stresses (e.g., deflection, buckling, and thrust by the pipe-soil system, traffic load, or pressure) acting on the pipe, and limit states of the pipe burst. Cluster 7 focuses on the estimation of stresses on buried pipes (Robert et al. 2022). Cluster 10 conducts structural reliability analysis using numerical simulation technologies (e.g., finite-element methods) to understand the pipe burst process (Hajali et al. 2016; Alzabeebee et al. 2018). Cluster 11 focuses on the construction of limit state functions (Rajani & Abdel-Akher 2012). Novel theories were introduced to measure the mechanical behaviour of stresses and establish burst limit states. Khemis et al. (2016) employed the subgrade reaction theory and first-order second-moment method to evaluate the effects of soil and structure properties on pipe critical buckling hoop force. Structural reliability analysis is useful to assess pipe burst risk, but the accuracy is limited due to the difficulty of accurate estimation of stress and pipe residual ultimate strength.

A pipe's physical condition deteriorates during its service time, and the deterioration process is influenced by various factors including corrosion, manufacturing techniques, pipe material, operating conditions, and environmental conditions (Rajani & Kleiner 2001). Basic mechanisms of the pipe deterioration process have been revealed such as corrosion growth, development of a crack, change in the pattern of pipe residual ultimate strength, and surface morphology of deteriorated metal pipe (Asadi & Melchers 2017; Wang et al. 2018). However, existing pipe deterioration models are not able to accurately simulate the actual complex process of a pipe burst. The pipe condition is difficult to monitor since WDSs are buried underground. Detailed information (e.g., pressure variation due to WDSs operation, temperature variation due to weather, and load variation due to traffic and soil) is not clear enough with existing monitoring technology. Then, the exact pipe physical condition is unavailable, which causes the parameter of laboratory experiments and numerical simulation not completely consistent with the actual situation.

Pipe burst can be detected by hardware-based, data-driven, and pressure wave methods. Hardware-based methods, including listening rods, leak correlators, ground-penetrating radar, etc., are widely and effectively used worldwide (Li et al. 2015). Data-driven and pressure wave methods are the research topics obtained by the reference co-citation analysis and are briefly introduced as follows.

Data-driven pipe burst detection methods explore normal patterns of hydraulic data (pressure and flow) and distinguish outliers to indicate pipe bursts. According to the analysis object, data-driven methods can be classified into one-point and multi-point methods. One-point methods focus on a specific monitoring site and identify outliers by exploring the time series pattern of hydraulic data (Misiunas et al. 2006; Mounce et al. 2010; Ye & Fenner 2011; Palau et al. 2012; Bakker et al. 2014; Romano et al. 2014; Laucelli et al. 2016; Wu et al. 2021; Zhang et al. 2022). Multi-point methods analyse numerical relationships of hydraulic data among different monitoring sites and detect abnormal hydraulic relationships based on their topological locations (Farley et al. 2013; Kang & Lansey 2014; Wu et al. 2016; Zhou et al. 2019).

Some typical data-driven pipe burst detection methods are summarized in Table 2, which were confirmed to be effective from validation results and played an increasingly significant role as water supply management tools (Mounce et al. 2017). Two technical points require special attention to apply data-driven detection methods. First, an appropriate algorithm is required for data analysis to pick up outliers from hydraulic data. Characters of hydraulic data of diverse WDSs are influenced by factors such as water use habits, network structures, water supply conditions, and so on. Then, selecting the algorithm fit for data pattern analysis is the key technical point for the establishment of a data-driven pipe burst detection method. The second technical point is to reduce false warnings of pipe bursts. Random water use and acquisition fluctuation of sensors cause fluctuations in monitoring data, which may be mistaken for a sign of a pipe burst, and reduce the accuracy and practical value of burst detection methods. Most pipe burst detection methods in Table 2 were applied to DMAs, the hydraulic data patterns of which were simple due to simple network typologies. The reliability of methods was enhanced by testing data mean variation during a particular time (e.g., from midnight to 4 am) (Romano et al. 2014) or defining burst thresholds in different dimensions (Zhang et al. 2022).

Data-driven methods can detect pipe bursts by online pressure/flow data analysis, which is of low cost and used to trace historical conditions compared to hardware-based and pressure wave methods. Under the extensive construction of online hydraulic data acquisition systems, novel algorithms for data pattern analysis are continuously led into this research topic. Data-driven pipe burst detection methods show good development and application future.

It is worth noting that attention should be paid to the applicable conditions of data-driven methods. Specific instructions on the requirements of monitoring points, monitoring data characteristics, and the structure and scale of water distribution system should be given. It is helpful to speed up the application of pipe burst detection methods and to give guidance for other researchers to improve the existing methods.

Pipe burst affects the transient trace of pressure wave, and additional reflections can be potentially used for pipe burst detection and localization (Chaudhry 2014; Rathnayaka et al. 2016). Colombo et al. (2009) summarized three types of methods for the identification and quantification of unusual pressure waves to indicate pipe bursts. The first was inverse-transient techniques that compared measured pressure waves between normal and burst conditions in the time domain. The second was frequency-domain techniques that utilized the changes in the frequency response of pressure waves to pipe bursts. The third was a direct transient analysis that aimed to isolate burst-induced unusual pressure waves in pressure trace. Pressure wave analysis is effective in detecting and locating pipe bursts, but has great limitations in practical use to real WDSs. The main reason is the huge investment in high-frequency pressure sensors.

The pipe burst process is complex due to randomly changing operation conditions (pressure, temperature, traffic load, etc.). However, it is feasible to predict the burst frequency of a pipe or burst rate of a WDS using appropriate prediction methods with project-acceptable accuracy. Five clusters (i.e., research topics) are obtained in this category as follows.

Pipe condition assessment is to evaluate if a pipe can perform its function through analysis of relative physical, environmental, and operational factors (Grigg 2006). Physical factors include pipe age, diameter, pipe material, pipe vintage, wall thickness, dissimilar metals, type of joints, pipe lining and coating, and manufacturing processes. Environmental factors include groundwater presence, soil type, soil moisture, climate, pipe bedding, pipe location, trench backfill materials, stray electrical currents, seismic activity, and underground disturbances. Operational factors include water quality, water pressure, backflow potential, leakage, flow velocity, and operational and maintenance practices. Various methods have been used to calculate factor weights, such as assignment (Alegre et al. 2006), ANN (Al-Barqawi & Zayed 2006), and the Bayesian theory (Wang et al. 2010). Pipe condition assessment is useful for maintaining acceptable levels of service and providing appropriate maintenance plans. The American Society of Civil Engineers determines grades of the national WDSs every 4 years according to the actual condition and required investment (ASCE 2017).

Relationships between factors and their relative contributions to pipe burst are distinctive for different pipe materials, buried geographic region, and operation conditions (Barton et al. 2019). It is impossible and unnecessary to use all these relative factors in order to develop pipe burst prediction models. Therefore, proper factors as input parameters of the prediction models need to be determined first. Asnaashari et al. (2009) found pipe length and buried depth were the main factors for bursts in ductile iron and polyethylene pipes, and pipe wall thickness was one of the main factors for bursts in ductile iron pipes. Wols et al. (2019) proposed that pipe burst rates were higher for cast iron pipes at low temperatures and asbestos-cement pipes at high temperatures. With the in-depth study of the pipe burst mechanism, factors like diameter, material, temperature, etc. are more and more considered to develop pipe burst prediction models, which is helpful for the improvement of burst prediction accuracy.

Machine learning methods automatically predict data tendencies using computers through experience (Jordan & Mitchell 2015). Breiman (2001) proposed the goal of machine learning was to improve prediction accuracy, which belongs to the algorithm modelling culture. Machine learning methods have advantages in pipe burst prediction owing to their self-learning to describe complex, distributed, and nonlinear systems and showed more reliable prediction accuracy or labor-saving than statistical pipe burst prediction models in some cases (e.g., Xu et al. 2011). Under the development of WDS information, construction, operation, physical, and environment data of WDSs have been greatly enriched, which provide preferential conditions of adequate and precise pipe burst record and factor data to apply machine learning methods. A large number of machine learning methods, such as genetic programming (Xu et al. 2011), Artificial Neural Network (ANN) (Tabesh et al. 2009; Jafar et al. 2010; Kutylowska 2015), boosted decision tree approach (Winkler et al. 2018), and evolutionary polynomial regression (EPR) (Berardi et al. 2008), are increasingly used to predict pipe burst in WDSs as efficient tools.

This kind of method skips mechanistic constraints and provides an effective and novel route for pipe burst research. However, they are highly dependent on monitoring data with noise, which need a burst mechanism to support model development (e.g., factor selection) and accuracy improvement.

Statistical methods can establish relationships between the pipe burst frequency (time) of an individual pipe or burst rate of a whole network and influencing factors. The statistical methods include Weibull distribution, Herz distribution, exponential regression, lognormal distribution, and time-linear regression, etc. (Scheidegger et al. 2015). They can be categorized into deterministic and probabilistic models. Deterministic models capture a definite burst pattern, and probabilistic models process pipe factor data to derive pipe burst likelihoods (Kleiner & Rajani 2001).

Statistical models have been applied to guide maintenance plans and future budgets as an effective burst control technique (Clair & Sinha 2012; Scheidegger et al. 2015). However, their accuracy is not very high generally, especially for individual pipes. Statistics derive empirical regularity of pipe burst frequency/time and make subsequent predictions based on the information of input factors. Model input factors cannot completely cover all influencing factors, since the existing monitoring system is incapable to collect the condition of all relative factors (e.g., pipe wall thickness). This demerit can be improved with the development of monitoring systems. Furthermore, there is the uncertainty of the pipe deterioration process due to random variations in operation and burial conditions. It is difficult to accurately predict pipe burst time even if the entire situation of a pipe is clear at present. This demerit can be improved with the introduction of novel methods to eliminate random variations. Consequently, there are common discrepancies between statistically derived predictions and the actual condition of an individual pipe. And burst prediction rate of a whole WDS usually has higher accuracy, since a large number of pipe samples can effectively reduce the impact of the randomness of an individual pipe sample. On the whole, pipe burst prediction using statistical methods is successful in the analysis of pipe burst trends, but there is still room for further improvement in accuracy.

Water utilities face the challenge of how to optimize the maintenance plan of WDSs with limited funds. Pipe burst prediction helps address this challenge by identifying specific pipes that have high burst risk and need maintenance in priority positions (Kleiner & Rajani 2010). A number of systems, including pipeline asset and risk management systems (Burn et al. 2003) and KANEW software (Herz 1998), have been developed to formulate maintenance plans based on pipe burst prediction. With the development of water supply industry, it is of more interest to establish multi-objective maintenance plans. In addition to the aim of pipe burst reduction, the objectives of water quality safety, reasonable budget, and energy reduction are also incorporated by determining inspection interval, pipe replacement subsequence, and appropriate maintenance technologies (Prosser et al. 2015; Romaniuk 2018). The establishment of a maintenance plan for a WDS becomes more systemic and incorporated.

Sensors collect hydraulic data of WDSs, forming the basis to comprehend the operation condition of WDSs. The design for use of sensors includes sensor installation position, construction convenience, sensor monitoring scope, investment, data storage convenience, and usable mode for data input and output. A certain number of research articles have studied the above areas for different WDSs in Cluster 5 (Simone et al. 2016; Qi et al. 2018). With the development of sensor manufacturing technologies and increasing requirements due to the construction of smart WDSs (Laramee et al. 2018), the depth and width of use of sensors continuously develop on how to use sensors to obtain precise and sufficient operating data in economical ways.

We probed the emerging trends of pipe burst research through five indicators, namely the evolution of the clustering network, evolution of each cluster, references with high citation rates, references with the strongest citation bursts, and references with high betweenness centrality.

One way is the emergence of new clusters (research topics), which indicates the development of research width. New emerging topics in pipe burst mechanism include physical conditions (Cluster 9) and limit state functions (Cluster 11), which help to understand and simulate the pipe burst processes. A new emerging topic in pipe burst detection is the pressure wave method (Cluster 12), which extends the research width of pipe burst detection. New emerging topics in pipe burst prediction include machining learning method (Cluster 1) and influencing factors of pipe bursts (Cluster 4), which improve prediction accuracy. A new emerging research category is the use of sensors (Cluster 5) to acquire hydraulic data. The research contents of those new emerging clusters mainly aim to improve the effectiveness and accuracy of burst simulation, prediction, and detection.

The other way is the cluster size increase, which indicates the continuous expansion of related research topics in research depth. Cluster 0 had the biggest size increase compared to the other clusters, which indicates the research depth expansion of data-driven pipe burst detection. The evolution of clustering network provides additional visual aids to identify changes between adjacent years in a flexible way, which helps new researchers to find the research hotspots.

All clusters show bright development trends indicated by the recently appearing nodes. For the clusters in pipe burst detection (Cluster 0) and pipe burst prediction (Clusters 1, 2, and 6), the research is deeper and more systematic indicated by more landmark references and co-citation relationships than other clusters. Other clusters (Clusters 3, 4, and 12) in those two categories are under sustainable development indicated by persistent landmark references and co-citation relationships over the timeline. The categories of pipe burst mechanism (Clusters 7, 8, 9, 10, and 11) and the use of sensors (Cluster 5) have a short history and weaker systematicness indicated by shorter horizontal axis length, less landmark references, and less co-citation links. But those clusters have attracted research attention recently indicated by their considerable co-citation relationships among nodes in recent years.

Table 3 lists 10 references with high citation rates of the dataset in our study. Seven references are in the category of pipe burst prediction (Berardi et al. 2008; Clair & Sinha 2012; Li & Mahmoodian 2013; Kabir et al. 2015; Scheidegger et al. 2015; Wilson et al. 2017; Winkler et al. 2018). Three references are in the category of pipe burst detection (Mounce et al. 2010; Ye & Fenner 2011; Romano et al. 2014). It indicates those two categories obtained more attention in pipe burst research.

Table 4 lists 20 references with the strongest citation bursts of the dataset in our study. Citation burst analysis whether the citation of a reference has statistically significant fluctuations within 1991 to 2022. References with the strongest citation bursts suggest the activeness of relating research topics and the values of their strength indicate burst intensity (Chen et al. 2010, 2012). From 2007 to 2013, nearly all the references with the strongest citation bursts were in the category of pipe burst prediction. Two references used statistical methods to predict pipe burst (Pelletier et al. 2003; Yamijala et al. 2009), three references used machine learning methods to predict pipe burst (Berardi et al. 2008; Tabesh et al. 2009; Jafar et al. 2010), two references were on maintenance based on pipe burst prediction (Davis et al. 2008; Kleiner & Rajani 2010), and one reference identified the predominant factors influencing asbestos-cement pipe burst (Hu & Hubble 2007). From 2014 to 2018, five references with the strongest citation bursts were in the category of pipe burst detection, which used the methods of the modified cumulative sum (Misiunas et al. 2006), transient-based detection (Colombo et al. 2009), principal-component analysis (Palau et al. 2012), artificial intelligence system (Mounce et al. 2010), and Kalman filtering (Ye & Fenner 2011) to distinguish abnormal hydraulic data. From 2019 to 2022, four references were in the category of pipe burst prediction (Wilson et al. 2017; Winkler et al. 2018; Sattar et al. 2019; Snider & McBean 2020). To sum up, pipe burst prediction has always been a research hotspot and pipe burst detection also obtained research focus recently.

The betweenness centrality of a node measures the importance of the position in the co-citation network, which is a measure relating to the transformative potential of a scientific contribution. Nodes with high betweenness centrality scores are always highly connected to other nodes or positioned between different clusters (Chen et al. 2012). Table 5 shows seven references with high betweenness centrality (no less than 0.10) of the dataset in our study. It indicates pipe burst prediction and pipe burst detection contribute more to establishing connections among the knowledge structure and tend to be more systemic in our defined area of pipe burst research.

This study shows that the knowledge structure of pipe burst research is classified into four categories as pipe burst mechanism, pipe burst detection, pipe burst prediction, and use of sensors. The entire pipe burst research is progressing remarkably, and computer science and technology show an important supporting role. The whole research tends to improve the effectiveness and accuracy of pipe burst prediction, detection, and simulation. Pipe burst mechanism and the use of sensors obtain remarkably increasing attention recently. Pipe burst detection has been attracting increasing attention with broad application prospects. Pipe burst prediction experiences deep and systematic research development, and provides evidence-based management suggestions for WDSs.

Reference co-citation analysis explores the evolution process of a discipline system from the perspective of citation relationships among articles, which is difficult to obtain through direct observations of researchers in a statistical manner. However, the process of discipline evolution is often characterized by multiple and heterogeneous components, complex structural relationships, and a variable and unpredictable external environment. Overdependence on co-citation analysis may lead to one-sided conclusions. Subjective understanding and judgement are still highly demanded for the analysis of knowledge structure and emerging trends.

This work was supported by the Ministry of Science and Technology of China (2019YFD1100105), the National Natural Science Foundation of China (52170105), and the Youth Innovation Promotion Association of the Chinese Academy of Sciences (2019043).

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.