By Jaime Valls Miro and Lei Shi, University of Technology Sydney
The Advanced Condition Assessment and Pipe Failure Prediction Project is coming up with a novel condition assessment research concept: exploiting data-driven research to improve large critical water mains condition prediction, over extended sections of pipeline, from limited condition assessment inspection data.
The Advanced Condition Assessment and Pipe Failure Prediction Project (ACAPFPP) is an innovative collaboration of researchers and water utilities from around the globe, dedicated to solving a major problem – failures in ageing critical pipelines which deliver fresh water to the towns and cities of the world.
Failures in critical pipes, generally exhibiting diameters greater than 300mm, present a major challenge to the cost effective management of water pipe assets. Targeting better renewals is essential to the success of the critical water main management framework.
Within the water industry, condition assessment (CA) prediction is mostly done by interpreting results between local excavation measurements using empirical and statistical estimation of extreme measurements (e.g. maximum pitting) with a considered high level of uncertainty.
Currently, the only feasible alternative is to use significantly more costly internal direct measurement tools, yet the effectiveness of this approach has not been fully established within the industry, given the significant deployment and validation investment.
The aim of this research activity is to develop a framework for ‘along the pipe’ condition assessment prediction with an improved level of certainty given current and upcoming practices for CA available to the water industry.
The current research strategy proposes the use of pipe condition data from targeted locally excavated inspections, possibly supplemented by additional screening information if and when reliably available.
This conceptual framework is currently being investigated with data from the critical pipe networks of the ACAPFPP utility partners, and the engagement from a range of commercial CA vendors.
Screening and local inspections
An example of the task to be addressed in this research is shown in Figure 1. On a buried pipeline, excavations are carried out at Sites 1 and 2, where the pipe is exposed, and local detailed CA inspections on these sites with tool A provide details of the pipe condition in the form of remaining wall thickness maps.
It is often the case that some limited information could be made available from other sources (for example, a screening CA tool). For illustrative purposes, in the case depicted in Figure 1, a screening tool B is shown, which is assumed to provide average thickness values, while an intelligent soil monitoring system C is assumed to be able to infer minimum thickness in the adjacent pipe section from environmental measurements (soil moisture, etc).
The objective of the framework proposed in this research is to predict likely pipe conditions in the form of a thickness map at unseen sections of this pipeline, based on inspection outcomes from Site 1 and 2, with or without extra information provided by screening system B and/or C.
Data correlation and machine learning
Data correlation plays a pivotal role in working with thickness maps and maximising the information gain from local inspections.
It is intuitive that the condition of a pipe at a certain location has some dependencies on its nearby locations’ condition.
We refer to such dependencies as spatial data correlation. The effect of data correlation is shown in Figure 2.
Describing data correlations in the condition of pipes in a one-size-fits-all manner is very difficult, maybe impossible. The main premise of this work is that this information, not readily available from traditional algorithmic mechanisms, can be better captured by data-driven machine learning methods, and can be exploited to reveal the condition of a pipeline between local inspection sites.
Machine learning is a powerful tool to uncover the hidden phenomena from the data directly. Specifically, probabilistic models able to capture data correlation can be learned from the available local inspections to shed light on inbetween predictions.
In the current work, non-parametric Gaussian Spatial Processes have been used.
‘Inbetween’ framework in practice
Assuming local inspection information from short pipe sections has been obtained, the overall method becomes a twostep process: modelling and sampling. The modelling phase extracts information from the specific local inspection data and generalises knowledge about the parameter of interest (e.g. wall thickness).
The sampling phase provides likely predictions by turning the generalised knowledge into probabilistically distributed samples, able to capture the underlying pipe data correlations at unseen locations.
These samples, usually in a large amount, collectively reflect reasonable variations in the expected condition of the pipe, providing that the learning has been generalised adequately.
Step 1: learning data correlations from local inspections, as shown in Figure 3. A data correlation model aims to capture and generalise the spatial dependency between thickness measurements.
Step 2: As seen in Figure 4, samples of likely pipes can be derived from the models learned in Step 1, without the need of any additional inputs.
However, given the information that increasingly available screening condition assessment tools are presumed to provide (e.g. assorted statistical values related to wall thickness), or more specific conditions from other indirect measurements (e.g. maximum expected pit depth from soil parameters such as soil resistivity), this can be exploited by data correlation models as a constraint to provide a refined prediction of the pipe condition.
An example of the kind of results that could be expected from this process is depicted in Figure 5, where all samples shown illustrate possible scenarios of a given area of interest that satisfy the measured constraint (e.g. the average remaining wall thickness as measured by a screening CA tool for that section of the pipe).
Given the probabilistic nature of the proposed datadriven methodology and the phenomena itself, exact inference of the condition of the pipe at arbitrary locations along the length of the pipe is not really feasible.
However the strength of the framework proposed in this research lies in its ability to derive likely conditions gathered from large numbers of sample representations.
This outcome format is beyond the state-of-the-art practice, and potentially supports a mechanism apt for subsequent structural stress or extreme thickness value based risk analysis, or any other reliant on the thickness semblance of a pipe.
The way forward
The aim of this research activity is to develop “along the pipe” condition assessment predictions with an accepted level of certainty by capturing the naturally occurring correlations in pipe wall thickness, given the constraints that inspecting large sections of critical water mains impose, and the perceived limitations of commercially available CA techniques.
The research concept provides a collection of statistically representative cases of likely “in-between” pipe wall maps which could ultimately aid in any subsequent analysis of failure modes and the remaining life of the asset. The model is currently under validation.