RAILWAY AGE AUGUST 2025 ISSUE: Artificial Intelligence is a controversial subject—though not for railroads when it comes to safety, if used properly.
The North American freight rail industry needs multiple technologies informed by artificial intelligence and machine learning to counter and address the complex needs of the vast North American rail network, sources told Railway Age. The railroads are responsible for monitoring 1.6 million railcars and more than 26,000 locomotives traveling on 140,000 of railway track located in diverse terrain, and AI-informed technologies ranging from ultrasounds and electromagnetics, to drone-based inspection systems, optical-based systems and LIDAR, help to do just that.
“When we talk about the railroads—and this is true for applications in other fields, like in aerospace or pipeline or any industry—not everything can be inspected, mainly because there is a lot of variability in the equipment part,” said Anish Poudel, a Principal Investigator II with MxV Rail’s nondestructive evaluation team. Poudel is involved in testing AI-informed technologies at MxV Rail. And that’s why “we rely on multiple technologies to kind of overcome the limitations of one technology versus another. So there is always a benefit of using multiple technologies that will allow you to see things differently,” he said.
AI’s Evolution in Freight Rail
The use of AI and machine learning in the freight rail industry has been well under way for quite some time. For instance, among the Class I railroads, western carrier BNSF uses AI in conducting wheel inspections, tracking container inventory at intermodal yards and making switching operations more efficient. Union Pacific has developed a ChatGPT-like tool that can analyze data trends.
Indeed, while AI has become a buzzword in the past several years, the concept has existed in the freight rail industry for decades, according to Poudel. In the 1970s and 1980s, the rail industry, along with other industries, utilized a low-level AI application known as the “perceptron technique” to develop an artificial neural network. However, the model ultimately failed because people were “under the notion that AI would solve everything—you build one model, and then you apply a wide range of applications,” Poudel said. “It was kind of like a one-size-fits-all kind of thing. And it failed.”
Despite the failure, people continued to build algorithms for AI, including those for machine learning, a subset of AI where a computer is taught to learn patterns. The advent of the deep learning neural network in the early 2000s opened endless opportunities for AI, including what people know today as large-language models, according to Poudel. “The railroads have been using some level of pattern recognition algorithm tools for many, many years. These were very low-level machine learning algorithms. But now, I would say that yes, the railroads do use advanced analytics,” such as deep learning models to process images from inspection portals that take pictures of speeding railcars, Poudel said. Such train inspection portals are in use at eastern Class I’s CSX and Norfolk Southern.
These inspection portals are comprised of multi-camera systems and lighting systems that can take millions of images of railcars going at track speed, according to Poudel. These images are fed into the deep learning models and enable the railroads to get real-time information on the health of the railcar, he said. The benefit of using this kind of technology is that it can enhance mechanical inspections, which can otherwise take an hour to inspect both sides of a 120-car train. The images can also be taken in the middle of the night, or if it’s raining or snowing.
Algorithms can also be developed to focus on searching for specific defects, such as broken wheels or defects on railcar springs. “You have an opportunity to do that inspection 24 hours a day, seven days a week, 365 days a year. So, in terms of productivity, I think using the machine vision systems, especially for looking into the rolling stock component, is a shift change,” Poudel said.
Indeed, one of the challenges in using AI now is the velocity and veracity at which data is coming out, according to Poudel. “This data is growing much, much faster than ever before,” and the data is imposing a lot of processing demands as a result, Poudel said.
One goal involving big data is whether it can be used to create real-time inspections, Poudel continued. Because so much data is being produced, “you have to rely on a machine to be able to assist you in terms of pinpointing where the problems are,” he said. This is why the human is still an important factor, because “you cannot just let the system go alone. You need to have the human in the loop there. There is a notion that, in different communities, the machine is going to take my job away, or something like that. It’s never going to happen. People are an important part of the equation.”
Poudel’s view of the human’s key role is also shared by Mika Majapuro, Vice President of Commercial Product management for Railinc. “Personally, I’m a big believer in humans and AI. It’s hard for me to see that AI would replace a ton of people in our field. You’re still going to need people who can look at the data and make critical decisions,” Majapuro said. “However, AI might support new employees in the freight rail industry by providing them with knowledge akin to that of an industry veteran. The big thing is, how can I make my new people perform like someone who’s been doing this for 20 years? And I think that’s where you’re going to see advances. And then, when there’s too much information, too much chaos, the tool can make some recommendations for you. It’s still up to the user to make critical decisions: What am I going to do with this prediction or with this recommendation?”
AI and Machine Learning Practical Applications
As Poudel mentioned, the freight rail industry has been using AI and machine learning for years. The products that are available now build more sophistication upon existing offerings.
Railinc has been working since 2018 on developing AI to improve the estimated time of arrival, according to Majapuro. Since Railinc’s initial offering of Advanced ETA, Railinc has produced at least three new updated editions. “We like to say that when we talk to shippers and ask them, what are your three top pain points? They say, ETA, ETA and ETA,” Majapuro said.
Railinc’s AI-informed product to improve ETA involves working with data provided by different railroad companies. For instance, NS may need to hand over a railcar to CSX, and Railinc functions as the middleman, where different data is passed between different parties with permission to view that data.
The product functions as a neural network model that looks at factors such as location, car type, commodity, time of year and time of the date to predict an estimated time of arrival. The data that feeds into Railinc’s ETA offering comes from wayside detectors along track in the U.S., Canada and Mexico. “As the car travels, we continue to make updates on the prediction,” Majapuro said. “Railinc’s ETA offering also uses historical weather patterns as an inpute, further enabling more informed decision-making.” For the freight rail industry, “sometimes it’s measured in weeks how long it takes to go from origin and destination, and you have crews that are timing out and trees falling on tracks. That’s why the ETA challenge has been so difficult for many companies to solve. There are so many variables. That’s why we’re so proud of the work that we’re doing, and being able to help the industry and help
the shippers.”
Another Railinc offering that uses data from the wayside detectors is one that monitors the wheelset health. This offering also uses neural networks, which “is obviously a different model than what we use for the ETA, but with some of the same principles,” Majapuro said. “You want to be able to replace those wheelsets before something bad happens. But also, there’s a convenience factor. If you can make predictions when the wheel set will fail, you can control when and how you’re going to change the wheelset.”
The wayside detectors have a reader that measures the force that the wheel is applying against the track, Majapuro continued. “We take that data, and then we take the time of the year, for example, and the miles on the current wheels, and we can make predictions on the likelihood that an alert will be triggered to replace the wheels.” The offering can also gauge the severity level of the alert, he said, allowing the railcar owner to maintain the car and replace the wheels when it’s convenient.
This type of insight is also valuable as a lot of very experienced railroaders are retiring, leaving behind a newer, incoming workforce that may lack institutional knowledge to make informed decisions. “You can use AI to make sense of that and prioritize tasks for you—maybe not trusting the AI to make the decisions, but the AI can propose to you, hey, it’s eight o’clock on Friday, these are the five things that you should be working on.”
Meanwhile, at MxV Rail, Poudel is involved in testing the application of machine learning on rail inspection that uses ultrasonic technology. The ultrasonic technology detects flaws at the microscopic level, and the machine learning will seek to characterize the flaws. Currently, the railroads collect data that eventually makes its way into the back office via an internet connection or some kind of communication architecture. The data gets processed using cloud architecture, and then the results get transferred to the maintenance group. (Download research paper below.)
But what Poudel and other researchers want to know is if machine learning can enable the railroads to make decisions on the fly, or as the data is being collected. “What we’re trying to do is, can we implement machine learning or AI in the same device where you collect the data and make that decision in real time,” Poudel said. “That would definitely improve the efficiency in terms of, you don’t need to go back to the track again and occupy the track and reverify it. If we can do that on the fly, that would be a huge step change.”
Poudel and MxV Rail researchers are also looking at applying machine learning using an edge computing device as well as saving raw data, which could open up the possibility of data fusion,” Poudel said. “The goal is the same towards the end: to prevent the rail failure,” he said. “So, our goal and vision is, can we come up with a methodology that would allow us to communicate with other forms of data stream and blend this data together, fuse this data, to come up with the high-level findings, and find or predict this discontinuities or anything that can go wrong with the rail ahead of the time?”
Collaboration is Key
Just as multiple AI-informed offerings and products are needed to respond to the diverse safety and operational challenges facing the freight railroads, so are multiple stakeholders needed to ensure that AI and machine learning are utilized effectively. “Input from these stakeholders may come from technical committees affiliated with the Association of American Railroads, or they may come from researchers’ white papers. Companies can take the findings from the resources and see how they apply to existing models and products, Majapuro said. “There are a lot of technical committees and meetings where our best practices are shared.”
Poudel agrees. In the committees involving the railroads, original equipment manufacturers and technology suppliers, “there is a great collaboration,” he said. “They don’t share exactly what they’re doing, but in general, they share experiences with each other that would allow them to work together and bring the technology forward. When you collaborate with different people, that’s where the innovation comes. So, when it comes down to AI, I think we as an industry need to figure out a way in terms of how we can share data and collaborate with others, because to be able to train a good model, you need tons and tons of data, and that’s only possible through collaboration.”
