It was exciting to read about Amgen’s AI-Assisted packaging inspection. According to the article “In an industry first, the biotech leader has equipped — and validated — an inspection system with artificial intelligence (AI) to boost particle detection 70%, cut false rejects 60%, and differentiate pesky bubbles from unacceptable contaminants in syringes.” (Bob Sperber, Packaging Digest 2/24/21)
I passed this along to one of my ‘go to guys’ for automatic inspection in the industry, Mike Guzik, to get his thoughts. His response was so insightful I felt compelled to share it with you.
Thank you to Mike Guzik!
Having worked with Syntegon (Formerly Bosch, formerly Eisai) in the past and continuing to do so, (currently validating a vial inspection machine). And since I have a background in computer science, I’m keenly interested to see how well this pans out.
In the last few years I have briefly had opportunity to discuss this with Dr. José Zanardi and his team as they targeted to develop this technology. Really he was looking for data sets that could be used as input test cases for the AI process they were developing. Very glad to see Jorge Delgado and the team at Amgen was able to help contribute to this effort.
As with all new technology, there is a learning curve (not sure if the pun is intended!), and some risks associated with early adoption. But if expectations and adoption is correct, then this should do two things:
- Improve the efficiency of the final process, in this case improve detection rate and reduce false rejects, sooner
- Improve on the time and effort it take to set-up the inspection process on the equipment.
To realize the first point, like anything else, good data is required. For these “AI-VI” systems as I call them, the more example images that the system can use for training, the more reliable it will be.
Additionally,
- These should cover the range of variability that the process will see, and
- This should be an ongoing process.
In order to really get an understanding of the full range of variability the process might see, inputs should be characterized over large data sets sampled under multiple conditions and this is best achieved over time. This means periodically augmenting the AI training with additional images to tune the AI network and re-validating as conditions and processes evolve.
The real power of AI is that it can sample data continuously and adapt to changing conditions continuously. However in the Pharma industry where change control is mandated, we should not be allowing the system to adapt and change on the fly unchecked. Reinforcement or refinement of the process by correcting what it got correct and what it got wrong during this process is needed to avoid drifting into incorrect thinking. Therefore the actual practice will be to collect images over time and periodically re-train the system, and re-test in a controlled manner, or at least to challenge the system regularly.
Regarding my second point, deployment, this will potentially allow more systems to be deployed as the vision engineering resource needed to set-up the system is at least partially performed by the AI system itself. All that is now needed is someone to present examples of good units and bad units (for each category) and to allow the system to program itself.
This means that if the engineering resources are a bottleneck to the deployment of these systems, more systems will get deployed faster without that bottleneck. On the flip side, if fewer engineering resources are needed for automated technology adoption, then this skill set, when actually needed to develop more AI systems, becomes relatively rarer and sough after as AI penetrates the market, increasing the divide between those who are able to create such systems and users who demand those systems.
The good news is that by automating the solution to known tasks, they can become more common (relatively more trivial), and the real entrepreneur can work on more high level abstract problems progressing mankind into the future!
Mike Guzik is President and founder of Robotic Automation & Machine Vision, Inc., and has worked with dozens of high tech manufacturing firms specializing in manufacturing startups, utilizing highly automated robotic and machine vision based systems.
Speak Your Mind