Kyle Murphy, Irth’s Director of Data and Architecture, recently shared with our User Summit attendees how Irth collects tremendous amounts of data that artificial intelligence (AI) and machine learning use to make valuable insights for the damage prevention industry.
In his presentation, Kyle challenged attendees to answer these two questions:
Q: In North America, do locate requests that indicate directional boring have a higher or lower chance of resulting in an incident?
Q: How much more likely is an incident?
A: A higher chance. Locate requests that indicate directional boring are 61.73% more likely to result in an incident than locate requests that do not indicate directional boring.
Did you get the correct answer? While attendees at the User Summit got the chance to win a Yeti tumbler by answering these and other questions correctly, artificial intelligence (AI) and machine learning analyzed the data to confirm the right answer.
This illustrates the insights made possible when artificial intelligence (AI) and machine learning analyze data.
The foundation of any predictive/machine learning system is twofold.
Irth’s data set includes the following:
Irth Insights with AI v1.1 is the culmination of years of data science investments. From the data science team with decades of experience in AI to acquisitions, Irth Insights with AI increases safety and saves money by empowering you to predict incidents before they occur.
There are two major elements:
Irth Insights with AI uses proven methods and predictive model training to predict risk, monitor and retrain predictive models when necessary, and ensure the delivery of the AI insights are available to as many people as possible.
Leading organizations, such as the Department of Defense, NASA, and insurance companies, quantify risk by considering the two components of risk: likelihood vs. impact. Irth quantifies and predicts risk using these proven methods.
Irth uses state-of-the-art AI on large data sets of highly relevant data that include national and targeted models. We started with a North American model that pulled a lot of data together to create the predictive models. We found through some of our studies that the game isn’t the same everywhere. For example, predicting incidents in Florida is very different than predicting incidents in New York, Minnesota, or California. Therefore, we took that reality into consideration to create targeted models.
It’s one thing to have really great predictive models, but how do we make sure they stay predictive and relevant amid drift, the systematic degradation of predictive models over time? Drift usually happens very gradually over a quarter or years, but it’s important to monitor it.
Irth has a lot of systems to monitor drift. For example, when explosives are involved in the excavation, we can monitor whether it’s still a predictive factor or is becoming less or more predictive over time. External factors can also contribute to drift, such as a New York law requiring excavator training or COVID. We can review drift to determine when we need to retrain our models or if external factors impact our predictions.
We have the most accurate risk prediction models in the industry. We protect you by monitoring drift, so we know exactly when and how to retrain models. This ensures learning from recent events and keeps predictions accurate over time, allowing Irth to remain the leader in risk prediction.
How we deliver these predictions is paramount. We want them available to as many people and as easily consumable as possible. Irth Insights with AI is fast, easy to integrate, scalable, and effective.
The Irth Insights with AI API is hosted on AWS via RESTful API, it’s highly scalable using the latest serverless architecture, and it’s optimized for performance, with a 6.4-second average response time. When emergency locate requests have a two-hour turnaround time, there’s no time to waste waiting minutes to get predictions and understand the risk.
Irth Insights with AI is an API, so while many people use UtiliSphere or DigTIx to get these predictions, it’s not required. The API is easily integrated with other systems.
Irth’s data science team is constantly improving our predictive models. In 2022, we completed a study to determine how seasonality and climate impact our ability to effectively predict incidents.
We found seasonality impacts predictive performance, so we updated the model training strategies to account for seasonality. Additionally, variation in precipitation from the average is not very impactful, but variation in temperature from the average is very impactful for low-seasonality states such as Texas, but less so for high-seasonality states such as Ohio.
Data science and machine learning allow Irth to identify and quantify what information predicts an incident. The predictive models we have trained have identified tens of thousands of trends and predictive factors.
At Irth, we use mountains of data and a team of dedicated professionals to uncover unlikely associations to help make informed decisions.
Midwest Energy, a gas and electric company, piloted Irth Insights with AI and a competitor. During the pilot, 100,000 locate requests were scored, no intervention such as an audit or watch and protect was performed during the pilot, and about 100 incidents occurred during the pilot. The results of the pilot were:
Not only is Irth Insights with AI great at identifying what locate request will result in an incident, it will also predict what locate requests will not result in an incident.
This is a huge result. Typically, if you can get predictions 1 or 2% better from AI and machine learning, it’s very impactful. Imagine if you could predict stock performance 1 or 2% better than a competitor. Using this predictive power is part of the solution to accomplishing a 50% reduction in damages in 5 years.
Irth is investing heavily in AI and our data science team. Here’s what’s on the horizon:
Here’s a real example of a locate request and the risk factors that were identified and generated by AI and a large language model:
The model predicted a high relative probability of an incident at 51.98% for the work ticket from Tennessee 811. This work involves underground electrical repair, as per the information provided by ACME Excavation.
The prediction was primarily driven by the nature of the work, specifically being a repair task, which increased the likelihood of an incident. Additionally, the number of members involved (12) and the area (3) also contributed to increasing the incident probability.
However, the fact that the request type was normal seemed to reduce the risk slightly.
The model’s confidence in this prediction falls between 44.16% and 71.49%, indicating a moderate to high degree of certainty in its assessment.
Irth’s skilled team of data scientists, engineers, and analysts have decades of experience with AI. Our goal is to empower you to predict incidents before they occur using the latest technology, making the predictions highly accessible, and explaining them in a way that is easy to understand and act upon.