The machines are watching: Quality control and defect detection with AI

Cameras can be placed ay various points in the production process to spot defects. (PHOTO: surasak petchang/Getty Images)

Years ago, quality control started off as a human being examining a product on an assembly line. Over time, we have progressed to machine vision, and more recently, to artificial intelligence (AI).

As promising, and as efficient as AI may be, according to Hugues Foltz, executive vice-president of Vooban, it’s not as widely used as it could be. “AI is not being used enough,” he says. “Instead, we see a lot of old vision systems that are not using AI at all.”

Foltz argues that older machine vision technologies are underperforming and are nowhere near as accurate as modern AI systems. Part of the problem, Foltz explains, is that AI is misunderstood, and is still a bit of a mystery. “The other part of the problem is that there aren’t a lot of companies that can implement these solutions,” he adds.

Moreover, AI is not a plug-and-play technology that can be easily implemented with out-of-the-box solutions. In order to effectively harness the power of AI, you first need a human being who can fine-tune the right algorithms.

Foltz says that these complex algorithms are the domain of AI scientists, which may be hard to find. “You need a certain level of expertise to understand and to fine-tune these algorithms,” he explains, “which is why an AI scientist is normally someone with at least a Master’s Degree in AI, along with a background in engineering, and that’s at a minimum. Ideally, you want someone with a PhD or a post-PhD.”

Fast implementation and equipment needs

While writing and fine-tuning advanced algorithms sounds complicated (and it is), according to Foltz, it doesn’t have to be time-consuming. In fact, he says that putting AI in place for quality control and defect detection doesn’t take long at all.

“It’s pretty fast,” Foltz says. “We can normally do a proof of concept within three or four weeks, maximum, if it’s something complex. The proof of concept demonstrates that we can see with accuracy what we are looking for, and in some simple use cases, that can even be done in a few days.”

Humera Malik, CEO of Canvass AI agrees, adding that data is a key factor. “Success depends on many things and starts with data,” she says. “However, in most cases, the data is already there. Companies have been gathering it, but not doing much with it until now. For any of our typical projects, going from training to implementation takes six to eight weeks. For success, ROI must be within immediate reach for any AI effort to be sustainable. For every dollar invested in industrial AI, we have seen an ROI of 10X and beyond, with operational efficiencies alone.”

As far as equipment is concerned, Foltz says that AI requires two basic elements: algorithms and cameras. “The equipment shouldn’t cost much,” he adds. “Cameras are common, they’re not expensive, and they already have a small processor in them that can run the algorithms.”

The fact that cameras are small and relatively inexpensive means that manufacturers can place them at various points in the production process in order to spot defects. In other words, rather than simply analyzing the final product, AI can be employed in conjunction with multiple cameras to spot defects at strategic points in the production process.

“We call this segmentation,” Foltz explains. “So, for example, we can have one camera looking at the shape of a part, another camera analyzing the colour later on the assembly line. This can save time and money because you’re not painting a part that should have been disposed of earlier in the production process because it’s the wrong shape.”

Today’s cameras and processors are so advanced, and so quick, that they can analyze an assembly line at incredible speeds, Foltz explains. “Speed doesn’t matter anymore,” he adds. “We have deployed AI in industries where the conveyor belts are moving so quickly that human eyes can’t see anything, yet the cameras catch 100 percent [of the defects].”

Perhaps the best news is that it doesn’t necessarily have to cost a lot to implement AI technologies for quality control and defect detection. Foltz says that depending on the size of the project, it can cost between $100,000 to $200,000 to deploy AI and train it to the point where everything works effectively and runs smoothly. “In contrast,” Foltz adds, ‘machine vision used to cost millions of dollars to implement.”

Canvass AI’s Humera Malik agrees. “Getting started can be very easy,” she says. “Typically, companies already collect an immense amount of operational data. The upfront cost commitment is minimal as there is no CAPEX investment. No change is required in the current resourcing and workflow. Most of the OT workforce is familiar with working with their data, so no new learning is required. One of the key factors is the organization’s commitment to learn from AI and implement the learning in order to see the ROI. In most cases, with data in hand, a typical project is implemented and shows results in six to eight weeks.”

As affordable as AI may seem, Malik believes manufacturers need to consider costs closely. “The cost of implementing AI may not be justified for all manufacturing processes, especially if the ROI is not clear or immediate,” she says. “While AI has the potential to revolutionize manufacturing, it’s important to assess its applicability on a case-by-case basis, considering the specific requirements and constraints of each application or process.”

Limitations

As sophisticated as AI may be, it does have limitations, according to Malik. “AI can be incredibly beneficial in manufacturing, but it does have limitations and may not be suitable for all applications or processes,” she explains. “Some key considerations include the complexity of the task. AI excels at repetitive and predictable tasks, but may struggle with complex, non-routine tasks that require human intuition and decision-making.”

In addition, she adds, AI systems require large amounts of data to learn from. “In processes where data is scarce or not digitized, AI may not perform well,” Malik explains. “In the industries we serve, which are heavily regulated, safety and compliance are paramount, and strict compliance is required, so AI systems must be thoroughly tested before being given full autonomous control.”

The future of AI

Foltz says that North American manufacturers are behind their European counterparts when it comes to the deployment of AI.

“If manufacturers still have people [on the assembly line] looking at things, they need to stop it, and they need to replace those humans with cameras and AI everywhere,” Foltz says. “I was in Germany recently, and it’s clear that automation in North America is lacking. [Europeans] are implementing AI on a massive scale and leveraging the power of automation. When they compete with us, they’re going to be cheaper because they are going to be faster. That’s something we should be worried about now.”

Malik sees a bright future for this cutting-edge technology. “AI will increasingly be used for quality control,” she concludes. “Today, it’s in the early stages. According to research firm, Gartner, by 2025, 50 percent of manufacturers will rely on AI-driven insights for quality control. That’s only half of the manufacturers, so there is a way to go.”

“In the industrial setting, we expect that AI will, in general, lead to more autonomous systems and further streamline processes. It will also advance to have enhanced predictive analytics capabilities and integrate more closely with IoT devices. So today, for example, the AI solution might be predicting something and alerting engineers when certain parameters veer from set conditions that would impact product quality. In the future, the AI solution could have complete control, and be able to take any necessary corrective action on its own.”