Prediction of pipe failure with Artificial Intelligence.

The knowledge exchange meetings on Hydroinformatics are organized in the framework of Waterwise, the joint research program of Dutch and Flemish water companies. An important goal of these meetings is to exchange practical experience and knowledge in thematic meetings.

Smart leak management

Steve Mounce (University of Sheffield, United Kingdom) started the meeting to point out the fast progression of data availability, evolving from his experience of just 5 district metered areas (DMAs) with 15 total data points on modem download in 1998 to 4,500 AMR measurements, collected every 15 minutes and downloaded hourly today in some cases.  Online sensors  including pervasive IoT sensing will bring about further progression. Based on flow and pressure sensors throughout the distribution network, commercial software using data driven analytics (such as ‘FlowSure’), enables utilities to locate and repair leaks faster. Steve overviewed a case study integrating additional pressure sensors, event detection and geospatial techniques for sensor optimisation and leak localisation within DMAs which was able to reduce the potential area for where leaks are occurring to 14%. This enabled faster localization and repair of leaks. Steve further showcased a risk hotspot map for pipe bursts. To better understand what causes or influences burst rates, the distribution system of the island of Barbados was taken as an example. Factors like pipe diameter, rainfall, month, season, year, land use, location, soil type, connection density, installation era, pipe material, and pressure were taken into account. In addition, shrinking and swelling of clay was found to be an important explanatory variable. A measure identified to decrease pipe burst rate is to increase bedding support for pipes. Pipe bursts themselves often lead to drinking water quality incidents. In the current ACQUIRE project, historic reports  describing incident details, responses and follow up are used to make a decision support tool to better handle water quality incidents, with help of Case Based Reasoning, Large Language models and Generative AI (AI assistants, Retrieval Augmented Generation, fine tuning, Agentic chains etc.). These generate high quality information in a ‘chatbot’ style interface and are being trialed by a UK water company in the control room.

AI may not always be necessary

Drinking water companies strive to get the best value out of their assets. Sometimes the assets can be a burden, if these do not function at their best. Jurjen den Besten (Spatial Insight) argued that the goal is leading when picking the solution, and not the use of AI per se. A cluster approach to reconstruct failure probability was shown to work very well, with future failure rate predicted almost perfect from the available dataset. With different methods available, Jurjen raised the question why it is hard to get methodologies embedded in the drinking water companies’ processes. One explanation in general is that change is slow in the drinking water utility practice. Employees that are new to the water sector may increase speed of adoption with their fresh view on things. This question was further elaborated on in the discussion at the end of the meeting.

Adding physics to models

Within Brabant Water,  a goal is to manage the number of failures per year in a way that is the most cost-effective. Dennis Schol, data scientist at Brabant Water, explained how ‘failure curves’ are constructed from data. The risk is expressed as risk per pipe length and material. By replacing pipes strategically based on the constructed model, the expectation is that failures will  stay under a desired level. By also taking into account direct and indirect costs of a possible failure and the cost of a replacement, an even more informed replacement plan is achieved. The system is currently used in practice to support asset managers that take the decision using their domain knowledge. Brabant Water is still aiming to improve the system by adding physics into the predictions.  For this goal, among others, the COMSIMA model of KWR,  describing pressure, pipe and connection degradation, is investigated. Other improvements are expected from adding information on construction date, batch numbers, physics-informed neural networks (PINNS) as a new technology, and LiDAR 3D data of the underground.

Demonstration of the ‘pressure tool’

The meeting continued with a demo presented by Mauk Westerman (Waterbedrijf Groningen). The presented tool, intended for leak detection, uses data from online pressure sensors in the distribution network. A challenge in finding anomalies is the high variability in pressure caused by water use. This is partly solved by normalizing the pressure for the pressure measured at connected pumping stations and considering the pressure at night when water use is low. This is fed into a machine learning algorithm using a sliding window of two weeks is considered to avoid feeding the model with current leaks. If the measured pressure is outside the predicted interval of the model, this is identified as a potential leak. In practice the tool has shown its worth by pointing out a significant leak in the distribution network that could have remained unnoticed otherwise, because it was not easily detectable by visual inspection.

A worldwide generalized model for pipe failure

Leave or replace, what is the best? Kevin Laven (Deloitte, Canada) presented results from his PhD research. The requirement to customize and train models per utility puts a large burden on the implementation of such models. Moreover, models for pipe failure will  generally yield better prediction results with a longer history of data collection and/or larger system. To circumvent this hurdle Kevin developed a generalized, ready-trained model based on data of utilities of all over the world. To have a model that is applicable across countries and industries, categorical data is generalized by the characteristics in the data instead of the named category itself.  The model Kevin developed provides two outputs for each pipe segment for an upcoming 5 year period: the probability of failure, and the expected the number of failures. The expectation value of failures can be used to calculate the expected number of failures within a cohort. These cohorts can be arbitrarily defined, however the utility likes, even after the model has been run. The resulting, generic model performed as well as any custom trained model. A 5.6 times improvement was achieved in preventing future breaks by replacing 10% of segments, compared to random replacement. The correlation coefficient between the expected number of failures and actual number of failures on cohorts with typical definitions (same material and diameter, same decade of installation, within the same city) was 0.99. An extra 15% improvement was possible when pooling data with those of other utilities.

Some way to go before implementation

Marco Dignum en Bas Jacobs (Waternet) presented their experience in applying the model of Kevin Laven to diagnose their pipe distribution network. The model can be applied directly (by running it on new data to predict failure frequencies) or indirectly (by using the results in other models, for example failure frequency curves in Rasmariant). Some uncertainty is involved in the application and several questions remain to be solved before definitive implementation. How reliable are the outcomes, do we need more checks? How many iterations should be done, are 5 years enough? Can we use it for cluster analysis? Can we couple it with GIS? When tested, the actual most failing pipes indeed fell in the higher likeness bins. The relative prediction is very good, but whether it is also good in an absolute sense is not (yet) investigated.

Adopting AI models for pipe failure

At the end of the meeting, the question of implementation is discussed. Several factors would enhance the real implementation of models. The procedures as normal will at fist be disrupted if a new model type is implemented, so this decision will have to be based on a lot of trust. What helps is if the engineer of the model is internal or well-connected to the utility. Also a good business case is necessary. The implementation should have a holistic approach; the decision to replace parts of the distribution system is usually dependent on many parties involved like for instance the city council.

Several methods for predicting pipe failure passed by in the knowledge exchange. All of them contribute to the better prediction of pipe failure. One of the main question for water utilities is which one to choose and what it would take to implement such models in practice, to replace old ways of pipe failure detection. A single model is, in any case, not going to dominate the decision system. Pipe failure models deserve more than being in a pilot stage, though.