How AI is Reshaping Manufacturing

Photo by Besjunior/Shutterstock

The use of automated quality control is becoming a must-have item in the toolkit of manufacturing businesses in order to prevail in the global competitiveness race. The application of such artificial intelligence-based solutions first gained widespread appeal in the developed world, where high wages made them a competitive alternative. By now, however, economies with lower wage levels are set to see similar proliferation due to the need to make up for a lack of skilled labor.

In Hungary, manufacturing is especially hit by labor shortage, and HR departments also face the problem of high fluctuation rates. According to a recent McKinsey study, automation, including smart quality assurance systems, can boost productivity by 30% as slashing costs helps rev up production volumes a great deal.

“Humans get exhausted after a while, and due to the repetitive nature of the work, their performance is also volatile. That’s the last thing industry needs,” stresses Attila Agod, owner and CEO of Machine Intelligence Zrt., an AI Coalition member company.

“Manufacturing requires predictability. Machines make the same decisions at the beginning as they do at the end of the shift, and although they also tend to make errors, at least they’re consistent even in terms of that,” he adds.

Apart from predictability, machines are also super-fast. Units can be assessed within 50-100 milliseconds, that’s how Heineken’s automated system checks some 80,000 bottles per hour with an error rate of 0%.

Reflection and shadows can cause problems, though; therefore, lighting and camera positioning are key factors to have excellent quality control results. More importantly, products and manufacturing processes change frequently.

Deep Learning Data Sets

“Because of the coronavirus outbreak, a Chinese supplier might need to be replaced by one from Thailand, but the material or the color of that new part might be different, and whereas a human operator would notice the difference right away, deep learning schemes need a huge data set to do so,” Agod highlights a current bottleneck.

Since products are evolving, models need to follow suit. Any change made in terms of packaging or manufacturing process means that the training set of the deep learning model has to be extended with samples from the new batch. Initial training can work with around 1,000 units; however, the training set needs to be constantly maintained and updated.

“This is where so-called active learning kicks in,” Agod adds. “This means that machines are going to be able to ask questions whenever it’s necessary. For instance, when it comes to labeling, the machine might ask a human operator for guidance whether in a certain case it’s OK to use any given label. Based on the response, its data set gets updated accordingly.”

What matters most is the variation of the acceptable products. Metal items tend to look similar, but pretzels can vary in shape, or for products with moving parts you need a large number of specimens for training.

Hungarian SME execs might be skeptical as to whether this is just mumbo-jumbo for behemoths with deep pockets, and not relevant to their world. Yet, reality is subtler than that.

“Our customers normally decide to invest in AI-driven quality control if the payback time doesn’t exceed two years. Company size is not necessarily what counts, though,” says Agod.

AI and Scaling

“We have customers with five and 50,000 employees. Size becomes a factor when it comes to scaling. If you have a production line that you want to duplicate, only the hardware costs double, the software needs to be developed just once. Accordingly, smaller companies should not fear to venture into AI quality assurance territory, especially if it is of critical importance in their processes.”

Machine Intelligence Zrt. mostly deals with foreign customers or the Hungarian subsidiaries of foreign companies. Substantial demand from Hungarian SMEs or multinationals is yet to gain on volume, but the interest clearly exists.

One of its customers is a startup that develops mobile apps to help eye doctors select the right glasses by optimizing the angle, position and size of lenses. The solution offered helps users precisely determine facial features based on a photo.

A German car manufacturer’s production line is another use-case: when putting together certain parts of the engine, pluggings need be to perfect, and it’s monotonous for a human to check in the long run. In this case, deep-learning-based on cameras come to the rescue. It is perhaps worth noting that auto makers in Hungary are all well-equipped with computer vision-based quality assurance systems.

“We clearly expect market demand to soar right after the coronavirus pandemic is over,” says Agod. “Decision makers understandably tend to put investment decisions on hold at times like these, but this crisis is certainly set to benefit our business in general, as companies will seek even more ways to engage automation.” Predictability of available workforce will play a more crucial role than ever, and AI-driven solutions can guarantee such predictability.

This is an area where many tasks formerly performed by humans can be taken over by machines. The human role is going to focus on supervising machines working on the basis of active learning. They help answer the “questions” of machines in case they are not certain of how to handle something. These algorithms also require constant development and training.

Annotation is yet another key factor to mention. Operators that used to observe units on the production line might be needed more to label specimens with the aim of improving data sets. So, it seems humans will not be replaced, but rather supplemented by the new tech, so that they become more productive.

Attila Agod