From wrinkle risk to critical barrier

Label wrinkles and dog ears on injection pens occasionally obscured key information like batch numbers and expiry dates. "If the wrinkles are in a critical area, like the expiry date, then it's an S3 error," explains the customer's lead technical specialist. "And if you have too many wrinkles in a batch, you might decide it's not acceptable to send it to the market."

A recent switch from paper to plastic labels only addressed part of the issue. "With plastic, at least the label doesn’t tear when you unfold it," he notes. "But that just pushes the problem to the patient — they're the ones who have to unfold the label to check if their medicine is still usable. That’s not good enough."

This wasn’t a new problem. For more than a decade, various teams had explored everything from cleaning adhesive residue off label rolls to experimenting with different label types. Some sites even tried using machine learning for image-based wrinkle detection.

But nothing stuck and the problem persisted.

The final push came in early 2023, when a deviation landed on the table. "That was the last straw," says the process supporter of handling it. "We’d had some before, but now it was necessary to find a solution."

"I fairly quickly landed on the idea of using a laser profiler," the lead technical specialist says. "It’s a bit like a 3D scanner." And the initial PoC confirmed that it was indeed possible to spot wrinkles this way.

Unlike traditional vision systems or AI models, the laser profile scanner measures height variations directly. "It’s just comparing numbers. You check if one number is higher than another. That makes it incredibly fast and easy to validate."

But what about machine learning?  "I’m a big fan of ML," he says. "But sometimes it’s overkill. Here, we solved it with a laser and a smart algorithm that can run on a laptop. That’s beautiful engineering."

Unlike machine learning, which tends to bring on severe validation headaches, this solution was elegantly simple. The sensor creates a 3D surface map of the label and spots wrinkles as height outliers.

"There's no guessing. It’s 100% measured," says the lead technical specialist. "That’s what makes it smart."

Better yet, the sensor is self-contained. No external light sources. No giant lenses. Just one unit. "That was a key parameter for us. It simplifies the setup enormously."

The CIM team joined early, helping shape the solution from the lab bench to the packaging line. Four sensors were installed across two high-speed lines, where each sensor had just milliseconds to deliver results.

"We usually get results in 50–60 milliseconds. If it takes more than 200, it’s a fail, simply because if it takes that long to measure, there are just too many wrinkles anyway."

The solution may be simple, but that doesn’t equal easy to make.

The technical solution had to deal with a tough mix of factors: high line speeds, vibrations, tricky lighting conditions, a multitude of label colors and the pen’s conical shape. And as the pens go into a tray after labelling, you can’t rotate them to inspect the full surface.

The collaboration was hands-on from day one. The lead technical specialist and Markus from CIM ran tests in the lab, on the line, and even on rainbow-colored labels to stress-test the sensor.

"We were very pragmatic. I tested vibrations, water, dust, label colors… We really pushed the limits. And Markus was always up for testing things out. He brought both theory and a practical approach to the table," says the lead technical specialist.

CIM delivered more than just the profile scanners. The SPECTS vision platform handles all data processing, as well as everything needed for compliance like user management and audit trails.

Because the system was retrofitted onto existing lines, the platform’s modular architecture was key. Only the newly developed modules — the wrinkle detection tool, the operator interface and the communication module for the sensor — require validation, making the implementation process a lot easier.

"I think technology is only 20% of it," says the lead technical specialist. "The rest is implementation, integration, and knowing how to work with surrounding systems. That’s where CIM really brought value."

The system is now live and collecting data on one of two lines.

There were some false rejects during the first few weeks, but the team has managed to fine-tune the system. "It's been running well for a month and a half," said the process supporter. "We’re in a really good place."

Despite real-world variation between batches, the system now for the most part runs without needing parameter adjustments between batches. "That’s a big win," he says. “It works for most scenarios and we’re testing what to do about the rare exceptions.”

The issue with wrinkled labels is well-known across different sites, so the goal is to scale the solution to other lines.

The technology could potentially be deployed on a number of lines that have been running for years, but since it’s a retrofit solution, it’s not just a copy/paste job. But algorithm-wise, there’s a lot that can be reused and that will make the process easier.

Expanding the use of this technology to other use cases is also a possibility. The lead technical specialist explains: "I tested it on open flaps. It could work there too. You can also use the same technology to check for things like buttons or caps. It’s actually quite versatile."

With the right mix of curiosity, pragmatism, and partnership, a persistent quality issue was turned into a scalable solution.

"We’re not just solving a problem here," the lead technical specialist concludes. "We’re building a platform for solving more problems in smarter ways. And that feels really good."