Understanding pipe failure is essential for effective asset management. Buried drinking water pipes are exposed to several types of external loads, e.g. soil weight, loads due to soil settling differences and traffic loads. The hypothesis that traffic loads positively affects the number of failures was statistically tested. For three out of four studied water companies significant higher failure frequencies than average were found at road crossings. Frequencies equal to average were found for pipes which are installed under other road sections. Frequencies higher than average–but not statistically significant–were found around speed bumps. The results of the multiple regression analyses show that the overall contribution of the parameter ‘road classification’ to pipe failure is small compared to the influence of pipe diameter, pipe material and year of installation.

INTRODUCTION

Understanding pipe failure is an important aspect in asset management of the drinking water infrastructure. Currently eight Dutch drinking water companies collect their failure data in USTORE; their uniform pipe failure database. In July 2015 USTORE consisted of more than 17,000 uniformly registered pipe failures, and almost 60,000 km of pipes. Analysis of USTORE data leads to a better understanding of drinking water pipe failures. In earlier studies on USTORE data relations were found between failure frequency and pipe properties such as, pipe material, diameter and age Kwakkel et al. (2013). These explanatory factors are all pipe related. Since buried drinking water pipes are exposed to several types of external loads through, e.g. soil settling differences and traffic loads there are also external explanatory factors. For these factors the (cor)relation with the pipe failure frequency is mostly unknown or not well studied yet. When drinking water pipes are exposed to external loads, the internal strain will be higher compared to a case without external loads, which may accelerate the process of pipe deterioration.

The current USTORE data format does not contain spatial pipe data on single pipe level. Without additional data, the current USTORE database can therefore not be used to find relations between failure frequency and external explanatory factors for pipe failure like traffic loads or soil settlement. This study was initiated to test to which extent traffic loads can be an explanatory factor for drinking water pipe failure. Failure data from USTORE was for this purpose combined with additional spatial data of drinking water pipes, spatial road data and geographical data of soil types.

METHODS

Two statistical approaches were used to test the hypothesis that there is a relation between traffic loads and the pipe failure frequency. The first approach focuses on specific locations with potential for peak traffic loads, namely road crossings and speed bumps. The second approach uses a multiple regression analyses to study the relation between the failure frequency of pipe cohorts and explanatory factors such as pipe age, traffic loads and soil type.

Failure frequencies at locations with potential for peak traffic loads

The first method focuses on specific locations were peak traffic locations could be expected. Two cases were studied: (i) road crossings and (ii) speed bumps.

Failure frequencies for these cases were calculated by the following four steps, which were performed within a GIS system:
  1. define location of interest (i.e. road crossings or speed bumps) as polygon data;

  2. calculate the pipe length in the under (1) defined polygons;

  3. calculate the number of failures in the under (1) defined polygons;

  4. calculate failure frequencies, based on pipe length, number of failures and the number of registration years using equation [1] below.

 
formula
1
wherein f equals the failure frequency, N the number of polygons, Fn the number of failures in the nth polygon, Ln the pipe length in the nth polygon and Rx the number of registration years for company x.
Polygon location data of road crossings was obtained by using the polygon road data from the Topographic Basis Registration of the Dutch governmental cadastral office (Figure 1). Polygon and point data of locations of speed bumps was obtained through the collaboration of nine Dutch municipalities.
Figure 1

Spatial data of drinking water pipes, pipe failures and roads in a GIS.

Figure 1

Spatial data of drinking water pipes, pipe failures and roads in a GIS.

Failure frequencies were calculated for three different types of spatial locations:

  • failure frequency at road crossings;

  • failure frequency at other road sections;

  • average frequency; the is the frequency calculated by all failures and pipes in the region of the company for which the frequency is calculated.

Multiple regression analysis

Several linear and non-linear multiple regression techniques were used to study the regression between the failure frequency of pipe cohorts and the explanatory variables:

  • Ordinary Least Squares;

  • Support Vector Machines;

  • Ensemble techniques: Random Forest, Extremely Randomized Trees Regression (ETR), Ada Boosting Regression and Gradient Boosting Regression (GBR).

The studied explanatory variables were (a) company, (b) pipe material, (c) pipe diameter, (d) year of installation, (e) soil type and (f) the road class (based on topographical road data obtained from the Dutch cadastral office). The road classification was used as a surrogate parameter for traffic load. Soil and road data was coupled to pipe data using a GIS. For each pipe attribute in the spatial database the following six steps were followed:

  • (a) determine drinking water company;

  • (b) define material (Asbestos-Cement, Cast Iron, Ductile Iron, PVC, PE, Concrete, Steel, others);

  • (c) define diameter class (1–150 mm, 151–300 mm, 300 + mm);

  • (d) define year of installation (1900–1919, 1920–1939, …, 2000–2019);

  • (e) define soil type (sand, clay, peat, urban area (i.e. unknown));

  • (f) define road classification when a pipe is buried beneath a road or installed close (within less than 0.5 m) of a road (main road, regional/local road, road within urban area of municipality).

After steps (a) to (f) pipe cohorts were defined by merging all pipes which share the same values for all attributes (a) to (f). Means (weighted to length) for diameter and year of installation were calculated for every pipe cohort. Finally pipe failures were assigned to the cohorts and the multiple regression analysis was carried out to find correlations between the failure on one hand and the explanatory variables material, diameter, year of installation, soil type and road class on the other. Predictions of the used regression techniques were compared with the real pipe cohort failure frequencies.

FAILURE AND PIPE DATA

Currently almost all Dutch drinking water companies share their pipe failures in USTORE, the uniform failure registration system of the Dutch drinking water sector. From USTORE failure data up to 5 registration years was available. The data subset used for this study contained almost 32,500 kilometres of pipe network and around 7,100 failures. The most commonly used materials are Asbestos-Cement, Cast Iron and PVC. The current USTORE data format does not contain spatial pipe data on single pipe level. Additional spatial pipe data was therefore provided by four Dutch drinking water companies. This data was geographically coupled to environmental data using a GIS. However, there was only spatial data available of the current drinking water pipe network. This shortcoming made it difficult to perform statistical studies on pipe material, since, e.g. a lot of Asbestos-Cement pipes are replaced by pipes of other materials during the last 5 years for which USTORE failure data is available. Since pipe failures were not linked to individual pipes this could lead to coupling of failures on removed pipes to newly installed pipes; which is obviously incorrect. The failure frequency analysis was therefore performed without distinction for different materials. For the multiple regression analysis data corrections were applied where older failure data did not match with current pipe data. These corrections prevented the coupling of ‘old’ failures to ‘new’ pipes.

RESULTS

Failure frequencies at road crossings

Failure frequencies at road crossings other road sections and average frequencies were calculated for four water companies (Figure 2). For three of the four studied companies significant higher failure frequencies were found at road crossings compared to other road sections and the company average frequency. Failures due to digging activities of third parties were excluded from the failure analysis.
Figure 2

Comparison of failure frequencies at crossings, roads and average frequencies. Failures due to digging activities of third parties are removed from the results. The bands show the 95% confidence interval based on a Poission distributed failure behaviour.

Figure 2

Comparison of failure frequencies at crossings, roads and average frequencies. Failures due to digging activities of third parties are removed from the results. The bands show the 95% confidence interval based on a Poission distributed failure behaviour.

Expert judgment lead to two hypotheses which could probably explain higher frequencies at crossings:

  • Higher failure frequencies at crossings are caused by higher vertical dynamic loadings at the road surface induced by accelerating traffic.

  • At road crossings often multiple (perpendicular) crossings of the pipe network and roads are found. Higher failure frequencies at road crossings are caused by point loadings which occur due to stiffness differences of road construction and the surrounding soil.

The results of company D show minor differences between the three calculated frequencies. This pattern is much different from the other companies A to C. This difference could possibly explained by the data acquisition of company D. A survey under the Dutch drinking water companies showed that companies used three different types of data acquisition to register the location (coordinates) of their pipe failures:

  • (A) acquisition of coordinates by the use of hand held GPS hardware in the field (companies A and C);

  • (B) acquisition of coordinates by the use of GPS hardware in the service vehicle of the fitter (company B);

  • (C) acquisition of coordinates by geocoding of registered addresses to coordinates (company D). In a second step these coordinates are corrected to the pipe position in the GIS.

The acquisition of coordinates according to option A may show the most accurate results.

Failure frequencies at speed bumps

Polygon and point location data of speed bumps was obtained from nine municipalities. The pipe length and pipe failures within the boundaries of these municipalities are a subset of the total database used for the study (56 failures; 335 km of pipe). Hence less failures were available to calculate failure frequencies at speed bumps, which lead to a lower statistical confidence. A combined failure frequency was calculated for speed bump locations in all municipalities to maximize the number of failures in the analysis and hence the statistical confidence. This frequency was compared to an average frequency for all municipalities, which was based on 697 failures and 4,630 km of pipe length (Figure 3).
Figure 3

Combined failure frequency around speed bumps in nine municipalities compared to the mean failure frequency of the drinking water network in these municipalities. The bands show the 95% confidence interval based on a Poission distributed failure behaviour.

Figure 3

Combined failure frequency around speed bumps in nine municipalities compared to the mean failure frequency of the drinking water network in these municipalities. The bands show the 95% confidence interval based on a Poission distributed failure behaviour.

As shown in Figure 3 the failure frequency at speed bump was slightly higher than the average in the study region. However, the frequency difference is smaller than the 95% confidence interval. Hence a significant increase at speed bumps was not found.

Multiple regression analysis in three drinking water networks

For reason of computation time the regression analysis was carried out for three of the four water companies mentioned above (A, C and D). In total 3,130 unique pipe cohorts were determined. To filter for statistically uncertain frequencies only cohorts consisting of more than five failures and more than 2 km of pipe length were used in the analysis. 195 pipe cohorts remained. Major materials were Asbestos-Cement (AC: 73 cohorts), PVC (71 cohorts), Cast Iron (CI: 34 cohorts) and PE (17 cohorts). In total 5,099 failures and 21,259 km of pipe length remained in all 195 cohorts for which the multiple regression analysis was carried out. Figure 4 shows the failure frequencies related to pipe material, installation year class, soil type, water company, diameter class and road class. The box-whiskerplots show the variability of the frequencies of all cohorts. The green star shows the mean frequency weighted over pipe length. The differences between the weighted mean frequencies for different road classes (0, 2, 3) are small. Pipe cohorts in peat soil show a weighted mean frequency which is approximately two times the frequency of other soil types. Other outcomes on pipe age, material and diameter confirm the outcomes of earlier studies performed on USTORE data.
Figure 4

Review of failure frequency per studied parameter. The box-whiskerplots show the variability of the frequencies for all cohorts. The green star shows the mean frequency weighted over pipe length.

Figure 4

Review of failure frequency per studied parameter. The box-whiskerplots show the variability of the frequencies for all cohorts. The green star shows the mean frequency weighted over pipe length.

The use of the non-linear ETR and GBR techniques gave the highest R2 value (0.6) in the comparison of predicted frequencies and real frequencies. For the results of these techniques the relative importance of the variable ‘traffic class’ was low; about 15%.

CONCLUSIONS

For three out of the four studied water companies significantly higher failure frequencies than average were found at road crossings. Frequencies equal to average were found for pipes which are installed under or close to other road sections. Frequencies higher than average–but not statistically significant–were found around speed bumps. The results of the multiple regression analyses show that the explanatory value of the parameter ‘road classification’ to the failure frequency of a pipe cohort is small compared to the influence of pipe diameter, pipe material and year of installation. This outcome complies with the outcome of the frequency analysis for road sections other than crossings. The results show that the contribution of roads to pipe failure is generally small. However, road crossings seem to deserve attention although the exact cause for the observed increase in frequencies at road crossings is still unknown. Understanding this cause will enable asset managers to take preventive measures. Moreover, from a risk perspective road crossings are locations where drinking water main failure could have great impact on above-ground and underground infrastructure.

Registration of the road class is currently part of the USTORE failure registration format. No general relation between pipe failure and the presence of roads (above or close to drinking water pipes) was found. Therefore it seems to be more useful to focus on specific road objects (e.g. speed bumps) instead of registering the road class.

REFERENCE

REFERENCE
Kwakkel
M.
Vloerbergh
I.
Van Thienen
P.
Beuken
R.
Wols
B. A.
Van Daal
K.
2013
Uniform failure registration: from data to knowledge
. In:
IWA-LESAM Conference
,
Sydney
,
Australia
.