HOUSTON — Tech startup Unspace was founded in 2020. Since 2022, we have been advancing machine learning in the field of machine vision to improve railway safety and operational efficiency, and our efforts are rapidly evolving. What began in 2022 as a concept to detect hot bearings using drones and machine vision soon evolved into the creation of a drone-based detection system and machine learning database for misaligned (bypass) couplers.
Phil Lentz, Unspace’s chief science and strategy officer, said Norfolk Southern wanted to identify hitch bypass situations that occur on-site when assembling car cuts into trains. A bypassed coupler involves a variety of scenarios, at least delaying trains leaving the depot, but it can also have more serious consequences. “The tension between the couplers when they are bypassed (due to a coupler failure) could result in a derailment on the premises.” [and] The key to the problem is that couplers can appear to be coupled when in fact they are not.
To solve this problem, Unspace aims to deploy drones equipped with high-resolution cameras to inspect freight cars in yards. A custom-tuned machine vision model is then applied to automatically analyze the visual data and identify bypassed couplers. Building the model turned out to be more complicated than expected, as Unspace needed data to help train a computer program to recognize coupling failure situations.
“Training data for machine vision models is important, but we didn’t have enough data in Southern Norfolk,” Lentz explains. NS didn’t have enough failures to work with. “So we partnered with Union Pacific to get more training data and also worked with BNSF,” he says. This allowed Unspace to establish a database of things like coupler bypass status.
Other challenges include teaching machines to accurately read reporting marks on rail vehicles, and the need to identify a specific vehicle and report its location when a problem is detected.
One idea quickly led to another. In creating the Bypass Coupler Database, Unspace developed the concept of an industry-wide database that can be used to develop machine learning models. This allowed industry participants to not only run the model, but also create their own models. Rail Genie is Unspace’s web portal platform for running machine learning models and sharing data.
How coupler bypass detection works
Character recognition to collect wagon report marks. Teaching a system how to read report marks has been difficult for a variety of reasons. unspaced
Data collection and analysis is the key to identifying failed couplers. An autonomous drone operating under an FAA Part 107 exemption (exemption for operating a drone beyond normal regulatory limits, such as flying beyond visual line of sight) flies at 16 miles per hour at a height of about 55 feet above the ground, surveying a cut in a vehicle assembled in a yard. The camera records high-definition video data up to the 4K standard (4,000 horizontal pixels per frame) at 30 frames per second. A machine vision model then analyzes the video footage.
Colby Bradley, vice president of rail innovation at Unspace, explains that there are situations that a drone’s high-resolution camera can visually detect from above, but that are “not necessarily obvious to a person walking along the tracks.” When viewed from the side or from the ground, the two joints appear to be touching each other, forming a solid connection. However, if you look from above, you can immediately see that the knuckles are butting up (the knuckles are closed and facing nose to nose) and are not connecting. This means that the vehicles are not coupled. “This is a failure that has to be corrected,” Bradley says.
Video quality has both benefits and challenges. said Tommy Davis, vice president of marketing.[When] We started the project and were really worried that the shadows would cause a big problem, that it would be too dark and we wouldn’t be able to see between the cars. Nighttime operations are still being improved, but daytime operations are not affected by shadows. [In situations where] Although difficult to see with the human eye, this technology overcomes contrast even at dawn and dusk. Computer vision can almost always see details better than the human eye. ”
However, it was difficult to clearly read the car’s reporting marks. Some vehicles, such as those carrying hazardous materials, have clear reporting marks stenciled on the top of the vehicle, but many vehicles only have reporting marks on the sides and ends. These should be captured as oblique images while the drone is flying alongside. The drone can adjust its flight angle to get a better line of sight. It’s common for cars to be covered in dirt and graffiti, making it difficult for cameras to read report markings. Bradley describes the difficult situation as follows: [recording] Technology using optical character recognition and AI. ”
Rail Genie provides important information to railroads in the form of reports. It identifies anomalies such as hitch or misalignment (wrong centering) and provides vehicle ID and location information so the railroad can act quickly on that information.
For large yards, especially the busiest hump yards, these reports allow railroads to be, in Lentz’s words, “prescriptive rather than reactive,” meaning they can identify and correct problems before they lead to delays or accidents. This is estimated to save the railroad up to $100,000 per day per yard.
Additionally, bypassed couplers can be quickly identified and resolved, increasing yard throughput and increasing efficiency by reducing dwell time delays.
new machine vision model
Trials of the Rail Genie coupler bypass model are underway at four hump yards (three in NS and one in UP), with plans to implement the technology in an additional 52 yards, many of which are flat-switched.
The success of this technology and advances through machine learning and data acquisition have opened the door to more applications.
“A classification yard can detect whether all cuts on a vehicle are safely joined together, as well as equipment failures that could create a hazardous condition for yard employees,” says Bradley.
Another advantage is that the computer can estimate how many times individual cars need to be switched before the entire cut is performed. As a result, the model can determine when the car is ready to pull the cut.
“This gives yard personnel a clearer view of the condition and readiness of vehicles in the yard,” says Bradley. “This is a very powerful tool.”
— To report news or errors, contact trainsnewswire@firecrown.com.
Share this article
