UNB News
News and stories from one of Canada’s top universities

Photo: Left to right. Joshua Pickard senior director of product and innovation at Eigen Innovations and Atah Nuh Mih UNB Master of Computer Science student.

The University of New Brunswick (UNB) Analytics Everywhere Lab in the Faculty of Computer Science has leveraged the power of AI to help a local company expand its range of products that detect defects in manufacturing sectors like automotive and paper processing.

Under the guidance of computer science professor Hung Cao, the Analytics Everywhere Lab (AE Lab) explored how to train a computer model to make decisions about manufacturing defects, based on data found in quality control images taken during the manufacturing process.

The joint project, called Automated Defect Detection Using Transfer Learning, brought together the leading-edge knowledge of AI capabilities from the AE Lab and the real-world industrial solutions being pioneered at Fredericton’s Eigen Innovations.

The company operates worldwide, providing products that help make manufacturing more efficient by analyzing data from machine vision systems and process sensors using intelligent algorithms to identify and predict defective products in real-time on an assembly line. Their systems have done everything from identifying coating defects on specialty paper using thermal imaging, to seeing critical surface defects on automotive parts using synchronized illumination and image capture, resulting in millions saved for manufacturers.

Eigen Innovations wanted to increase the applications its technology could be used for, and the collaboration with the AE Lab on the project began.

“[Eigen Innovations] built a platform, and it worked well in certain scenarios, but it’s really challenging to scale it up in certain applications,” said Cao.

Identifying defects in manufacturing conditions that produced images at different angles or lighting conditions or from different products posed a challenge. That’s where the AE Lab and Cao’s expertise in AI modelling, Internet of Things, and data analytics came in.

“Once you introduce unwanted variations into the dataset, model performance degrades,” said Joshua Pickard, UNB alum (BScE'12, MScE'14, PhD'18) and senior director of product and innovation at Eigen Innovations. “The work that Dr. Cao had done was to help find a way around performance degradation.”

The solution, provided with the assistance of AI, was to take a very “localized” view of the defects.

“A defect, when you zoom in, it looks the same on these different parts. But when you zoom out, it’s very different,” said Pickard.

By focusing on the similarities, the model has an easier time identifying the defects. The hope is that the work that Cao and the AE Lab have done will eventually be packaged into an Eigen Innovations product.

“This is phase one for our initial collaboration,” said Cao of the work of the AE Lab, which was funded by a New Brunswick Innovation Fund (NBIF) AI Pre-voucher Fund.

“We have the result, we have the prototype, we have the proof of concept, we have the paper published,” he said, noting that there could be more collaboration on the fruits of this project with Eigen Innovations in the future.

“Our focus has always been: if you build it the first time and you solve it the first time, you should be able to scale it to new deployments,” said Pickard. “This research was all tied to that. We want to be able to solve it and make it generalizable so that we can scale it to other deployments, other machines, other applications, where the data is similar enough that we can get sufficient results.”

Computer science master’s student Atah Nuh Mih, who worked on the project with Cao at the Analytics Everywhere Lab and the Eigen Innovations team, agrees.

“We were able to develop a processing method that could really get a huge data set just from a few images,” said Mih. He says their approach allows the model to be applicable to different objects without having to retrain it to recognize new things.

Eventually, he hopes the model will be refined to the point where the training can be done right on the manufacturing floor, in the plant itself, without doing everything on the cloud.

Mih says the big advantage of having a model that can operate independently on the manufacturing floor is privacy and protection of business processes.

The other advantage is versatility.

“When you have training done at the manufacturing plant itself…you can easily and quickly retrain the model if you want, instead of taking all the data to the server and redoing everything and deploying the model again,” he said.

For Mih, the experience of getting to work with industry partners was valuable.

“Working with industry, you get that ‘hands-on’ feeling,” he said. “It’s something that really matters to the world…You’re contributing to something that can really help people and make an impact.”

Mechanical engineering professor Rickey Dubay is also a co-author on the project’s research paper and Dubay himself has a long history with Eigen Innovations.

Eigen executive chair Scott Everett (BScE'09, MScE'12) co-founded the company with Dubay after completing his master’s degree under Dubay’s supervision. The product Everett created became the basis of Eigen’s business and the collaboration between the two continues today.

Many Eigen employees are former UNB graduate and post-graduate students, and act as honorary research associates, maintaining a strong relationship between industry and the university.

“We now co-supervise MScs and PhDs to fulfill the projects that are on Eigen’s roadmap,” said Dubay.

Dubay says keeping the “knowledge base” in New Brunswick was a goal of the ongoing collaboration between Eigen Innovations and UNB.

“That was our mission, to train highly qualified personnel, to hire them and to keep them here,” he said. “The local high talent is still here and it snowballs into other [industries].”

As Eigen Innovations has grown, so has UNB’s expertise in manufacturing defect detection – a niche market that has the potential for collaboration among a number of different academic disciplines.

“I don’t know everything,” said Dubay with a grin. “I’m specialized in my field. And that’s where the multidisciplinary approach of inviting profs in different departments comes in.”

For Cao, the experience AE Lab computer science students like Mih get from working on real-life projects with industry partners, especially those that run across disciplines, is important.

“The skills they learn from doing multidisciplinary research collaborations with companies and other academic partners is one of the best outcomes of the project.”