Vision Week banner

Create Scalable Vision and AI Solutions with a Systems-Level Approach to Data

June 10, 2021
Advances in the data science community are enabling smaller manufacturers to take advantage of AI, ML and advanced machine vision.

As the manufacturing industry pushes deeper and deeper into Industry 4.0, the requirement for advanced machine vision and AI will become standard. A lot of things have to happen for these technologies to take root. Simplified end-to-end data solutions, quick defect data labeling and taking data scientists out of the equation can help smaller companies partake and create their own factory of the future.

A3’s Vision Week hosted panelists from Intel, Landing AI, Integro Technologies and NVIDIA who discussed all these issues and how the factory of the future will be driven by vision.

David Austin, senior principal engineer at Intel explained companies that are searching for insights from their data are moving the market forward.

“What we’re starting to see is a lot of scale coming to market,” he said. “There are solutions you can now get from systems integrators, there are solutions you can get from open source.”

It doesn’t stop at insights, though. Austin said there is more to be done with process control and autonomy in the future.

Piyush Modi, business & development chief strategist at NVIDIA, described three major themes he’s seen in the past four years:

  • Industrial inspection becoming ubiquitous
  • Cameras armed with superhuman vision capabilities in distributed lines
  • Far-reaching digital twin technology

“Machine vision is and always has been an enabler of digital transformation,” added David Dechow, principal vision systems architect at Integro Technologies. “Machine vision has been used in industrial settings for 50-plus years.”

As data tools and algorithms improve, they open the doors for machine vision even wider. But different goals require different data amounts and data requirements.

“There’s semi-supervised learning that says ‘hey, we don’t need all of this labeled data. We can use some labeled data and some unlabeled data.’” Dechow explained. “If we’re smart about the data we collect and smart about what we label, we don’t need gobs and gobs of data.”

Daniel Bibireata, VP of engineering at Landing AI, built on Dechow’s data argument and said that it really doesn’t take a tremendous amount of data to create models.

“We have found that you can get very good models if you have as low as 20 productive images per defect class,” he said. “The unambiguous and accurate labeling of your defect data—we found that to be extremely important.”

So, how do we get from data labeling and algorithms to the idea of the factory of the future? The personnel have to get involved.

“Training used to be the painful thing in terms of making it efficient for these data scientists, who are really far and few,” said Modi. This drives up demand for the scientists and, in turn, hinders smaller companies from adopting the technology that calls for them.

“We gotta start removing data scientists from the loop,” Austin, a data scientist himself, said. “We’ve gotta start enabling some subject matter experts to do more and more of this.”

How do we get this technology into the hands of smaller companies? “Making [deep data analytics] easier and easier is what’s going to move that factory of the future away from the huge players and into the medium and small players,” said Dechow.

Pre-trained algorithms and data toolkits, like NVIDIA TAO, can give data science to companies that don’t have the resources for data science personnel. This allows OT personnel to get involved.

“They can participate in this journey and put their knowledge and they can see firsthand how the food is getting cooked,” Modi said. “So, gone are the days when they say there is no trust between the providers and the users.”

Since companies are all on different pages of their factory of the future journey, and have very different use cases for advanced machine vision technology, it can be difficult to come up with a solution and collect the data used to create a model.

“Instead of going for days, weeks, months trying to gather datasets to train, we’ll implement a discrete analysis that works right out of the box,” said Dechow. “But the data outgoing is the bigger issue.”

As the technology advances, anomaly detection models are getting quicker and cheaper, and as Austin points out, can even be implemented in a plant in which machines are not connected to an IT infrastructure.

“I see AI and data science, along with machine vision, going in the same path,” he said. “We’ve got to simplify things, and then the low-hanging fruit really starts to come to reality.”

Sponsored Recommendations

How BASF turns data into savings

May 7, 2024
BASF continuously monitors the health of 63 substation assets — with Schneider’s Service Bureau and EcoStruxure™ Asset Advisor. ►Learn More: https://www.schn...

Agile design thinking: A key to operation-level digital transformation acceleration

May 7, 2024
Digital transformation, aided by agile design thinking, can reduce obstacles to change. Learn about 3 steps that can guide success.

Can new digital medium voltage circuit breakers help facilities reduce their carbon footprint?

May 7, 2024
Find out how facility managers can easily monitor energy usage to create a sustainable, decarbonized environment using digital MV circuit breakers.

The Digital Thread: End-to-End Data-Driven Manufacturing

May 1, 2024
Creating a Digital Thread by harnessing end-to-end manufacturing data is providing unprecedented opportunities to create efficiencies in the world of manufacturing.

Voice your opinion!

To join the conversation, and become an exclusive member of Machine Design, create an account today!