For years now, artificial intelligence (AI) has been working behind the scenes to improve the everyday lives of people. Based on the data it collects, AI can recommend what to watch next
on Netflix and can suggest spellings for that word that we just can’t get right. It also lets us know when to take a different route home because of a car accident or bad weather conditions and even knows when to flag abnormal credit card spending to stop cyber thieves in their tracks.
In the future, AI will be enhanced to facilitate autonomous cars and will be able to gauge the emotions of students through facial expressions to determine areas where they may be struggling. It will also be leveraged to diagnose diseases in patients faster and with more accuracy. And, as has already been seen in the manufacturing industry, AI-powered robots will take on the dangerous and monotonous tasks that humans would prefer not to do.
An exciting example of that can be seen in Novarc Technologies’ NovEye weld monitoring and control software. Powered by AI, NovEye is a powerful system that helps arc welding robots adapt to welding and perform defect detection in real time.
Not only does it improve quality and productivity, it has the potential to bring more individuals into the manufacturing industry. The technology is of more interest to younger generations that might not be inclined to do the physically demanding work welding jobs traditionally require. It also creates opportunities for technicians, developers and engineers to join the industry and continue to improve the way AI is utilized in manufacturing operations around the world.
Machine intelligence
Just like any other non-manufacturing AI system, NovEye’s ability to monitor and make real-time corrections to the welding operation is based on years of data collection and machine learning. Soroush Karimzadeh, co-founder and CEO of Novarc, explains how the technology arrived at where it is today.
“The machine learning models have been trained to dissect the images of a weld and are essentially told what the needs are,” he says. “There’s an adaptation process of teaching the machine learning model about the features of the weld, as in ‘this is the side wall, this is the center of the root, these are the edges of the bevel.’ Once the machine learning model is trained, you test its performance and perform a validation to the other side. After that, it goes into implementation.”
Taking a step back, these machine learning models rely on artificial neural networks, which are computing systems inspired by the human brain that serve as the backbone to the powerful algorithms that drive machine learning. Like a child learning to walk, AI learns by trial and error, adding additional layers to the algorithm as time passes. As such, the AI can be trained by data to understand what an object is – if a picture of an animal looks like a cat or dog, or in the case of NovEye, what the center of the welding root looks like.
“Another exciting area within the concept of neural networks is transfer learning,” Karimzadeh says. “Think of somebody that has learned to play a violin; technically, they can learn how to play a piano because they now understand the notes or how basic music works.
“If the NovEye technology knows how to weld a fillet weld on pipe, for example, it’s thanks to transfer learning technology and this emerging body of work that we can transfer that knowledge into how it can weld a fillet on flat material,” he adds. “Each new model, therefore, doesn’t require to be trained from scratch.”
Cobot control
Currently, NovEye works in tandem with the company’s Spool Welding Robot (SWR), which was introduced to the industry as the first pipe welding collaborative robot or cobot, for short.
As referenced in its name, NovEye is the vision-based software that increases the autonomy of the cobot, further freeing the operator to work on other tasks.
“Human welders use their eyes and their biological neural network to command their hands to do the great welds they do, and for robots to do anything similar, they would have to follow the same path,” Karimzadeh says. “So that’s why NovEye is a real-time, vision-based system. It has to be able to analyze the weld pool and what’s going on with it, just like any welder would be able to do.
“We designed and trained it to basically design certain features in the weld dimension the same way welders would be looking at the weld pool and how it’s wetting on the sides, the position of the torch and then deciding what to do next, whether it’s increasing or decreasing weld parameters, wire feed speed or travel speed, or adjusting the many other parameters that are at their disposal,” he explains. “The neural network or AI would be doing the same thing. It’s been trained to look for those features, and it’s been trained to command the control system to make those corrections during the weld.”
Understandably, a lot of work and a capable team was needed to produce a vision system that would be able to control a weld pool with the performance that’s required, in terms of latency and for and weld quality. Karimzadeh says the system had to be implemented in a low-level code so that it can run efficiently on the robot, “and that took some time for us to do because it requires a considerable amount of effort to implement high-quality, reliable code at those low levels..
“Today, NovEye is a big giant piece of software that ties into the control system of the robot, and that’s so the robots – the actuators – can control the weld puddle the same way human welders would,” he says.
Partner productivity
That giant piece of software and all of the data that was required to develop it is in part thanks to the beta users with which Novarc has been working. “We couldn’t have done it without having our Spool Welding Robot out there doing the welds and generating a ton of data,” Karimzadeh says.
Novarc is based out of Vancouver, so one of the company’s primary educational partners is the British Columbia Institute of Technology, which has multiple faculties, including its welding school. Novarc has also worked with Kwantlen Polytechnic and other trade schools in the Vancouver area.
In terms of SWR customers, Karimzadeh says there are, in fact, a few that currently have the NovEye software and are ramping up their use of it.
“Once NovEye is fully rolled out, the operators set up the SWR, select the recipe they want to use, turn on the controllers, hit start and the process of the weld starts,” he says. “Three seconds after that point – with NovEye – the operators don’t have to do anything else. The machine does the weld from root to cap, adjusting for every scenario that is required. Without it, the operator has to watch out for the roots, change travel speed or correct the torch position, if it’s required. So, there’s still basic intervention required from operators. With NovEye, they don’t need to do any of that.”
The ultimate goal, he adds, is getting to the point where an apprentice can set up the SWR and let it weld up joints, “let’s say 10-in. pipes for 10, 20 or 25 min., and just let it run.” While the cobot is doing the work, these operators could be prepping another weld or even a second or third SWR, increasing the productivity of these shops beyond the productivity enhancements they achieved when initially adopting the cobot.
“Don’t get me wrong, though, we still need welders,” Karimzadeh concludes. “We need great welders to do the welds that no robot can do at this point. But the boring, repetitive welds that are dangerous? They’re not very rewarding, and they don’t require the higher skill of experienced welders – those can be delegated to the intelligent robots.”
