Artificial Intelligence (AI) and Machine Learning (ML) are terms we are hearing more in the last few years. The roots of these technologies are varied and include the discovery of modern statistical methods from the 1800s; neural networks, the theory of which was created in the 1940s; digital computers from 1950s onwards and most recently, access to high-performance computing and large file storage. Enabled by these technologies, a recent abbreviated history of AI & ML goes a bit like this. In 1997 Gary Kasparov lost at chess to IBM’s supercomputer Deep Blue. In 2016 Lee Sedol lost to AlphaGo at the game of Go, draughts like a board game with many decisions to make. Around the same time, in 2012, AlexNet won the ImageNet competition by recognizing images with an accuracy of around 85%. With these three advances together, computers were now officially able to perceive and make decisions. When coupled with electric motors, hydraulics, and other actuation technologies, we can now have a walking, talking, thinking robot. Both positive and negative views of the future are available and being worked on by academics and experts. Either way, we now have capabilities that we did not possess before to deploy.
In today’s factory, AI and ML are being promoted as solutions for product quality improvements, cost reduction, and workforce skills improvement. This promotion is relatively recent and linked to the reduced cost and increased access to computing power that these technologies need to function. Using a convolutional neural network (CNN) similar to AlexNext, SUNY University at Buffalo and Moog have been able to recognize the difference between good and bad parts produced via additive manufacturing. This is exciting stuff and has the potential to free up inspection resources or provide new quality metrics at minimal additional cost. However, an inspection of AM components remains a single element of solving for the industrialization of Additive Manufacturing. As with other changes in manufacturing we at Moog are certain that adoption and use of this technology will come back to practical application and return on investment. The next example explores this further.
"AI and ML are being promoted as solutions for product quality improvements, cost reduction, and workforce skills improvement"
For centuries, manufacturing has been focused on processes efficiency. As we enter the latest generation of improvements, using machine vision with machine learning technologies, we face the challenges of finding the right applications where value is created, developing the right talent to best take advantage of ML & AI and understanding the limitations. Some businesses make the wait and see approach by watching others develop solutions to adapt. Others actively invest in learning to build a differentiator for themselves. Both plans come with great benefits and risks. We believe in a hybrid approach is best by investing in technology in concert with collaborating with others to minimize the risk. This approach also helps accelerate the learning because you will have multiple organizations (peers, universities, OEM’s) bringing alternate perspectives.
Once an organizational approach is determined, what are the potential benefits and limitations of AI/ML? As an example, operator inspection consistency using a microscope to view excellent features can be tough to maintain. Current best practice includes regular training, detailed work instructions, and frequent breaks to avoid loss of concentration or physical strain on the human body. Since computers can see via digital cameras and perceive via AI, inspection processes are prime opportunities to increase throughput, increase repeatability, and reduce operator fatigue. By adding machine-learning capability, computers will also be able to improve the ability of the inspection process with increased precision over time as more data is gathered.
On the surface, this use case for machine learning looks to be good. However, occasional process extremes can be hard to accommodate. The defect type you are trying to detect is typically bounded by the data that the machine was trained with. If you prepare the system to discover discoloration in part and then a different or new surface defect appears, this new or modified feature will most likely escape the automated inspection process. The machine was never trained to look for the surface defect and will often not be able to ask a question. Human inspectors have more general intelligence and cope better with edge cases. Consequently, they are more capable of noticing unknown defects and ask questions if they see something different or strange. Training the AI to look for these new surface defect is possible, but it will require an updated algorithm.
A return on investment for machine learning applications also depends on the availability of a different type of skill set in your organization for the creation and support of machine learning algorithms. The person or team of people at the sharp end of implementation will need statistical math skills, software development skills, along with a deep understanding of the operational application. Due to challenges of edge cases, we expect operators to be assisted by AI/ML vision systems rather than being wholly displaced. Human operators working together with a computer based AI/ ML system will be better able to cover all the unusual conditions that the manufacturing process can deliver. The discoloration/surface defect scenario described above where an operator and AI work together is an example of this. (In the world of chess this combination of human and computer based AI has proven to be very powerful and has been named Centaur Chess or Cyborg Chess) Following an AI/ML implementation if the operator is retained and specialized skills were needed for the application, a more extended term payback is likely. There is no doubt that Machine Learning and Artificial Intelligence both have their place in the future of manufacturing, but the overnight transformation of the manufacturing workforce and the processes that they guide appears unlikely.