Artificial intelligence and machine learning can diffuse water’s ticking time bomb

What makes artificial intelligence better at solving water infrastructure challenges?

Cities are full of sensors. But what are we doing with all of that information? What if the data could tell us that we are leading to a failure? What if it told us we could extend the life of an asset, help prevent failure, and save an expensive and extensive repair?

A key factor in improving the efficiency of managing water systems is to implement an integrated approach to asset management. Machine learning is the process of building an intuitive model that “learns” from past similar or related cases without requiring specialist intervention. Let’s face it, I love my job as an engineer and number cruncher, but computers do this type of task best.

This learning seeks to identify underlying structures or patterns that are meaningful in understanding the often-subtle relationships between the water line breaks and a variety of factors. Those can include age, soil condition, break history, valve operability data, hydraulic model data, land use, and parcel data.

Linking data to artificial intelligence (AI) allows business and technology to evolve together. Programmed correctly, these systems can predict, learn, create, and relate. AI increases the scale of services, thereby delivering more work and improving productivity. AI can exceed our expectations and human limitations.

According to a study from Price Waterhouse Coopers, AI will contribute up to $15.7 trillion to the global economy by 2030. It’s estimated that 45% of total economic gains will come from AI-assisted product development and stimulating consumer demand. There are many opportunities for the useful application of AI in the water industry. The advantage of AI in speed, analytical capacity, and efficiency is a game changer for utilities evaluating risk. But most importantly, it will help meet future societal needs.

Using AI to build your water asset risk curve

Water main breaks are a nuisance to the public. But to utility owners, they can have potentially severe consequences. Litigation, bankruptcy, and skyrocketing costs are all possible.

Let’s look at prestressed concrete cylinder pipe (PCCP), a common variety of large-diameter pipe used for transporting water and wastewater. PCCP, like other pipe materials, is eventually going to fail. Predicting the failure of PCCP was traditionally developed using risk curves based on Finite Element simulation. Since 2010, I have provided technical assistance and support for the development of more than 1,400 performance curves worldwide. That work has saved utilities more than $10 billion (estimated) in repair and/or replacement costs. And now, I’ve automated it.

Selection of the right risk curve model is significant. It is the core functionality that learns from past data and generalizes into the future. Such models have been used for different tasks, including classification, regression, anomaly detection, and synthesis. We need to take four steps for this methodology to work.

After data sources and collection have happened, we can determine if the data quantity and quality is sufficient for forecasting. The second step—feature importance—is an element of my proprietary platform that assigns a quantitative score to available data (pipe material, age, soil condition, amount of overburden, break history, and pressure data). Then it determines which factors are most dominant in affecting the water breaks. Identifying the most dominant parameters in water breaks provides essential information for the development of the criticality study, reduces unexpected failures, and increases our ability to develop a successful predictive maintenance plan.

Step 3—model selection—is especially significant as models can been used for different tasks, including classification, regression, anomaly detection, and synthesis. In step 4—validation—an evaluation is used to determine the confidence level of our predictions. The final design is presented as a curve with a level of confidence regarding its accuracy.

Taking information overload and turning it on its head

The traditional method for understanding the performance of PCCP requires technical expertise and is a costly and time-consuming process. Typically, it requires two to four weeks to generate a performance curve. Because of high liability and high technical expertise’s requirements, only a handful of firms offer these services. The condition assessment software tool I’ve discussed turns several weeks into seconds.

The application of machine learning-based algorithms to develop improved performance curves/failure risk assessment achieves better maintenance strategies based on the learned information and improves the accuracy and efficiency of asset management. It also informs lifecycle cost analyses, proactive operations/maintenance programs, and capital replacement plans based on sound financial analyses, such as cost-benefit and net present value, within integrated data control and acquisition systems.

If underinvestment in our infrastructure continues, the nation’s water systems will become less reliable. Breaks and failures will become more common. We will see disruptions compound, and the nation’s public health and the economy will be at risk. This type of condition assessment through machine learning can reverse this trend.

With these advanced data driven processes, we can predict water main failures and effectively manage aging conveyance infrastructure, reduce unexpected failures, and increase the ability to provide clean and safe drinking water. This means less headaches, uncertainty, and lower costs to rate payers.

Want to learn more about how this digital approach to advanced asset management practices can change the trajectory of infrastructure operation and maintenance? Reach out, let’s talk.