Earl Yardley, director at Industrial Vision Systems, writes about the recent impact of automation on the factory floor.
Integrated quality inspection processes continue to make a significant contribution to medical device manufacturing production, including the provision of automated inspection capabilities as part of real-time quality control procedures. Long before COVID-19, medical device manufacturers were rapidly transforming their factory floors by leveraging technologies such as artificial intelligence (AI), machine vision, robotics, and deep learning.
These investments have enabled them to continue to produce critical and high-demand products during these current times, even ramping up production to help address the pandemic. Medical device manufacturers must be lean, with high-speeds, and an ability to switch product variants quickly and easily, all validated to ‘Good Automated Manufacturing Practice’ (GAMP). Most medical device production processes involve some degree of vision inspection, generally due to either validation requirements or speed constraints (a human operator will not keep up with the speed of production). Therefore, it is critical that these systems are robust, easy-to-understand and seamlessly integrate within the production control and factory information system.
Deep learning
Historically, such vision systems have used traditional machine vision algorithms to complete some everyday tasks: such as device measurement, surface inspection, label reading and component verification. Now, new “deep learning” algorithms are available to provide an ability for the vision system to “learn”, based on samples shown to the system – thus allowing the quality control process to mirror how an operator learns the process. So, these two systems differ: the traditional system being a descriptive analysis, and the new deep learning systems based on predictive analytics.
Innovative machine and deep learning processes ensure more robust recognition rates. Medical device manufacturers can benefit from enhanced levels of automation. Deep learning algorithms use classifiers, allowing image classification, object detection and segmentation at a higher speed. It also results in greater productivity, reliable identification, allocation, and handling of a broader range of objects such as blister packs, moulds and seals. By enhancing the quality and precision of deployed machine vision systems, this adds a welcome layer of reassurance for manufacturers operating within this in-demand space.
Deep learning has other uses in medical device manufacturing too. As AI relies on a variety of methods, including machine learning and deep learning, to observe patterns found in data, deep learning is a subfield of machine learning that mimics the neural networks in the human brain by creating an artificial neural network (ANN). Like the human brain solving a problem, the software takes inputs, processes them, and generates an output. Not only can it help identify defects, but it can, as an example, help identify missing components from a medical set. Additionally, deep learning can often classify the type of defect, enabling closed-loop process control.
Deep learning can undoubtedly improve quality control in the medical device industry by providing consistent results across lines, shifts, and factories. It can reduce labour costs through high-speed automated inspection. It can help manufacturers avoid costly recalls and resolve product issues, ultimately protecting the health and safety of those towards the end of the chain.
AI limitations
However, deep learning is not a silver bullet for all medical device and pharmaceutical vision inspection applications. It may be challenging to adopt in some applications due to the Food and Drugs Administration (FDA)/GAMP rules relating to validation.
The main issue is the limited ability to validate such systems. As the vision inspection solution utilising AI algorithms needs sample data, both good and bad samples – it makes validating the process extremely difficult, where quantitative data is required. Traditional machine vision will provide specific outputs relating to measurements, grey levels, feature extraction, counts etc. which are generally used for validating a process. With deep learning, the only output is “pass” or “fail”.
This is a limiting capability of deep learning enabled machine vision solutions – the user has to accept the decision provided by the AI tool blindly, providing no detailed explanation for the choice. In this context, the vision inspection application should be reviewed in advance, to see if AI is applicable and appropriate for such a solution.
Conclusion
In conclusion, deep learning for machine vision in industrial quality control is now widely available. Nevertheless, each application must be reviewed in detail – to understand if the most appropriate solution is to utilise traditional machine vision with quantifiable metrics or the use of deep-learning with its decision based on the data pool provided. As AI and deep learning systems continue to develop for vision system applications, we will see more novel ways of adapting the solutions to replace traditional image processing techniques.
