Trueflaw takes a crack at world dominance
Train axle inspections could prove to be a lucrative business for Trueflaw with its new technology.Credits: : Pexels/anna-m.w.
A world-first application of machine learning is set to break Trueflaw onto the world stage.
If artificial intelligence does become self-aware at some point in the future, we’ll likely be thanking Iikka Virkkunen and his team at Trueflaw. Their new machine learning-based system will most likely be busy inspecting train axles, nuclear power plants and space vehicles for a living, seven days a week beyond the accuracy of seasoned human professionals.
Until the emergence of artificial consciousness, though, Trueflaw will in any case have their hands full with the solution’s current capabilities. The new technology is able to accurately spot microscopical cracks in train axles and cut around 75–90 per cent of the tedious work conducted presently by humans in non-destructive testing (NDT) of various materials. And not to worry, the AI has been designed to support professionals, not replace them. Yet.
“Our new technology can save human analysts time by doing the bulk NDT analysis for them,” elaborates CEO Virkkunen. “The professional can then focus on the more important tasks which require human judgment, enhancing decision-making.”
Virkkunen estimates that the solution’s market is worth one billion euros and will continue to expand. The first pilot for the machine learning technology was completed recently, demonstrating that something deemed impossible just a few years ago is now in the hands of a three-person company in the outskirts of Espoo.
“I usually tell people this is the third impossible thing we’ve done as a company,” smiles Virkkunen.
“It can’t be done”
To understand the new technology better, it is crucial to go back some 15 years to the beginning, to the Helsinki University of Technology that is Aalto University today. Virkkunen and his fellow co-founder, Mika Kemppainen, were finishing up their doctoral theses in thermal fatigue research or, in simpler terms, breaking stuff with heat.
“If you think about an accident like the malfunctioning of a machine, the last bang you hear is usually the end of a long fracture process,” explains Virkkunen. “The existence of cracks is a marginal occurrence, though. On the other hand, if there are cracks, they’re potentially of catastrophic consequence.”
Around the same time, the nuclear industry had developed a need to improve reliability testing and the duo began to receive visitors examining their test data and research.
“We quickly realised we had developed a lucrative technology to produce controlled natural cracks,” tells Virkkunen. “We patented the idea and have been doing it for the past 15 years although our first to-be-clients assured us that producing and managing the cracks is an impossible feat.”
No one is questioning the team any more, at least not in the industry.
“My daughter was asked in kindergarten what her parents do. She told them that her dad has a crack factory,” chuckles Virkkunen. “We had a discussion with the kindergarten personnel after that remark and had to explain daddy’s occupation a bit more.”
“Yeah, we’ve tried it, doesn’t work”
The majority of NDT is currently mechanised, meaning that NDT professionals first scan the material in question with different techniques and then sit through hours of data identifying fractures.
“That’s really the challenge in the inspector’s job. You’re looking at data day after day and what you’re looking for is rare but crucial. You can go 10 years inspecting these things and not find a single crack.”
Mechanised testing produces digital data that is used in training new professionals, which carries with it promises for machine learning. However, the data was limited for a long time due to the cost of creating more of it.
“Especially in the nuclear industry, the training specimen are just like the real things that can cost up to hundreds of thousands of dollars and weigh tens of tonnes. So, the analysts have a chronic shortage of training material which could be used for training and development.”
Or had until Trueflaw’s second impossible feat.
“In 2014, we developed a technology to manipulate the digital data produced by mechanical inspections,” says Virkkunen. “We are able to digitally alter the flaws found in the inspections, to add stuff and take things out and even create new cracks into the inspection data.”
The technology is known as eFlaw and was essentially an answer to the cost-efficiency question. It also enabled the generation of vast amounts of data.
Trueflaw’s ability to produce raw data (cracks) and manipulate it digitally into relevant quantities (eFlaw) led the team to a simple conclusion.
“We’re teaching this stuff to humans, so why not machines as well?” says Virkkunen. “We’re not machine learning geniuses, we just happen to have the best data and the means to create more of it.”
The technology is still quite new but is charging ahead with real momentum. With several demos lined up for 2020, the Trueflaw team proved their solution in a recent pilot with Dekra, a 3.3 billion-euro vehicle inspection company, and Helsinki City Transport (HKL), using it successfully to inspect subway trains’ axles in Helsinki.
“We were able to cut the inspection time of a train’s axles from four hours to an hour,” says Virkkunen.
The team are confident about the geographical and cross-industrial scalability of the concept.
“If it works in Helsinki, it should work in Mumbai as well. We’ve also had talks with nuclear power plants, for which we have a pilot starting next year. The potential is basically everywhere mechanical inspections are made.”
“We’ve been doing this for 15 years and just happened to find ourselves as the only ones in the world that can make machine learning work in this niche industry,” says Virkkunen. “Call it an informed surprise.