The defect has already happened by the time inspection finds it

On a wallboard production line, a sheet can look perfectly sound even as a void begins developing beneath its surface. The problem may not become visible until the material is cut and inspected further downstream, by which time the conditions responsible for the defect may have affected considerably more production. Yet subtle changes in the material’s thermal profile could have provided an earlier warning, while there was still an opportunity to intervene.

This is where the distinction between inspection AI and manufacturing AI becomes interesting. One can determine whether a finished product meets specification, while the other can help identify the conditions likely to cause failure before the product is completed. Erin Barrett, President and CEO of Eigen Innovations, believes manufacturers are overlooking the value of those earlier signals.

“We find most problems have early indicators,” Barrett says. “Usually, there are things happening in the process, process drift that you can pick up before actual costs or defects hit the operations. The sooner you can get that problem identified and fixed, obviously, the better your P&L looks and the more competitive advantage you have.”

The cost of finding out too late

Inspection performs an essential containment function but often sits downstream of the operation responsible for the defect. A worn cutting tool may damage several components before sampling identifies a dimensional problem, while molding conditions can deteriorate before the finished part reveals it. The later a fault emerges, the harder it becomes to establish how much production was affected.

To understand whether a defect is about to emerge, manufacturers must connect current process conditions with the quality of parts produced later. Machine settings, temperatures and inspection outcomes need to be associated with the relevant product and production stage. Engineers can then investigate whether certain patterns signal genuine deterioration or merely normal operating variation. The problem is especially difficult on mixed-product lines, where a temperature or pressure pattern that is acceptable for one product may signal trouble for another. Material batches, recipes and tool identities must travel with the measurement if engineers are to distinguish a meaningful relationship from coincidence.

Eigen’s work in extrusion blow molding illustrates the opportunity. Defects in a finished product may originate while the hot parison is forming, before conventional sampling reveals a problem. Thermal imaging can show changing conditions at that earlier stage, when operators still have an opportunity to intervene. “Instead of just getting a value of the temperature of a product or an asset, you’re actually getting a continuous stream of thermal images, and you get a story versus a point in time,” Barrett says. “The system continuously learns and adapts, so it’s looking for things that look unique and flagging when something out of the norm happens.”

Not every unusual temperature pattern predicts a defective part, and factories cannot stop whenever a measurement changes. Engineers must establish which warning signs consistently precede unacceptable output. Only then can an alert justify intervention rather than create additional production disruption.

AI does not make conventional inspection obsolete

A scratch rarely looks exactly like the one before it, and contamination can be equally unpredictable, changing in appearance depending on the material and lighting conditions. These variations make conventional inspection rules difficult to apply consistently, creating opportunities for deep learning to identify features that fixed algorithms might miss. However, recognizing a potential defect is only part of the task. Manufacturers still need reliable measurements and tolerances to determine whether the product is acceptable.

Christian Eckstein, Product Manager for the Deep Learning Tool at MVTec, argues that effective inspection combines techniques rather than assuming AI should replace everything already in use. “One customer told me that they do not have a single AI vision solution that does not also use traditional machine vision,” Eckstein says. “There is almost always some preprocessing or post-processing. The AI may segment the defect, but afterwards you still must measure it and decide whether its size, shape or other characteristics are acceptable.”

A pharmaceutical inspection system at Aspen’s Notre-Dame-de-Bondeville facility in France demonstrates the hybrid approach. It captures up to 14 images of each ampoule, combining deep learning for foreign-particle detection with conventional vision checks. Requiring suspected particles to appear across multiple images helps reduce false positives and unnecessary rejection of acceptable products.

A straightforward dimensional check may need only a deterministic rule, while variable defects can justify machine learning. Production cycle time must still govern technology choices. “You can have a model with two or three percent more accuracy, but if it takes minutes to process one image it is absolutely impossible to use in an industrial setting,” Eckstein says. “You have to select models that work in your environment, so there is always a trade-off.”

Make inspection results explain what happened

The evidence of a manufacturing defect often lies several stages upstream from where it is eventually discovered. A component may fail its final inspection because of a temperature fluctuation, an incorrect machine setting or a variation in raw material introduced much earlier in production. Without a connection between inspection results and the conditions under which the product was manufactured, engineers can identify what went wrong but struggle to establish why.

Erin Barrett of Eigen Innovations describes how this problem can develop in wallboard manufacturing, where small pockets of air can become larger voids that are only discovered during downstream inspection. Thermal monitoring offers a way to recognize subtle changes while the material is still being formed, allowing operators to investigate before the defect develops further. By combining those images with process records and product-specific thresholds, manufacturers can distinguish normal variation from conditions that warrant intervention.

An inspection system that repeatedly rejects perfectly good products can become almost as costly as the defects it was installed to detect. Beyond unnecessary scrap and rework, constant false alarms can erode operator confidence until warnings are routinely questioned or ignored. Manufacturers therefore need to calibrate detection thresholds around the characteristics of each product and the consequences of missing a genuine defect, rather than pursuing maximum sensitivity at any cost.

The information collected during inspection can also provide value long after a product has left the line. When images are linked to operating conditions and individual production records, manufacturers gain a detailed history of how each component was made. Engineers can use that evidence to investigate recurring failures, while quality teams can retrieve the relevant records when dealing with customer claims or suspect batches. Investigations that previously required piecing together information from disconnected systems can instead begin with a traceable account of what happened during production.

Earlier signals need a route to action

Recognizing a developing quality problem changes nothing if the evidence never reaches the person or system able to intervene. A technician might inspect a tool, or an operator might adjust a setting within a validated process window. More serious conditions may justify an automatic response, but manufacturers must establish which decisions remain with experienced production personnel.

Barrett describes applications where defined severity levels allow AI-generated signals to instruct a PLC to remove a product or stop a line, while less serious warnings remain with operators. “We’re pretty keen on pushing for automation where it’s possible,” Barrett says. “In a lot of our applications we have warning signals that don’t do an automatic thing without a human in the loop. But if it gets to a certain level of severity, we’re sending a trigger to a PLC to either remove the product from the line or stop the line.”

Stephen Graham, Vice President Product & Technology for Hexagon’s Production Software division, approaches prevention from the engineering stage. He sees value in AI-assisted CAM programming before a machine starts cutting material, when some production risks can still be addressed digitally. On a live line, the corresponding opportunity is to act on credible process evidence before more defective components are made.

Earlier intervention does not justify uncontrolled autonomy. A model may identify a correlation without proving that its recommended adjustment will solve the problem. “If it gets it wrong once, you can cost hundreds of thousands of dollars on the shop floor,” Graham says. “For anything that’s really going to make a difference, you’re always going to want that verification piece. I doubt we’ll get to the point where we fully delegate responsibility on the shop floor to AI.”

Turning prediction into routine quality control

A warning system that performs well during commissioning may become less dependable as products, equipment and operating conditions change. Eckstein has seen the difficulty of transferring an inspection model between factories. “Customers often set up a production line in one country as a pilot and then scale it to other sites, and it does not simply work in the same way,” he says. “The camera may be slightly different; the lighting may be different and even the parts may vary more than expected.” Local feedback and controlled retraining therefore become continuing engineering responsibilities as production conditions evolve.

A sensible starting point is a recurring defect that carries a measurable cost. Engineers can test whether early indicators remain useful across batches, then monitor whether warnings actually prevent failures. First-pass yield, scrap and response time reveal more about the benefit than model accuracy alone.

None of these developments removes the need for reliable inspection because not every defect can be predicted in advance. The opportunity is to use what inspection reveals to improve the upstream process, giving production teams a chance to act before more value is lost. Rejecting bad parts protects customers, but preventing their creation improves manufacturing performance.