Pores, spatter, bubbles: The welding robots in the automotive industry can get any number of things wrong. And they have plenty of opportunities to do so: Up to 100 welds are required before a piece of sheet metal can be called a car door, for example.
Robots complete the welded joints in fractions of a second – far too quickly for the human eye to register or even control the process. Until now, the automotive industry has relied on manual quality control, which requires a good eye. Sometimes, companies use computer tomography to assess the quality of the weld or even choose the path of destruction and separate individual weld seams again to detect any defects. Whatever methods the car companies choose, these kinds of subsequent checks are always expensive, time-consuming, and don’t cover everything.
In a joint research project with a major German car manufacturer, T-Systems has shown that Lenin's assertion “trust is good, control is better” can be turned on its head with the help of artificial intelligence. The prerequisite is that the robots used in the production are able to develop an understanding of the quality of their work. It’s only then that carmakers can trust their laser or spot-welding machines to detect their own mistakes, immediately sound an automated alarm in the event of deviations, and remove the defective parts. So many conditions cry out for the use of artificial intelligence: That's why T-Systems is raising the IQ of welding robots with a combination of deep learning and reinforcement learning.
“Cameras and sensors provide the necessary information about the welding process, ”, explains Jörg Heizmann, Senior Sales Manager Big Data & AI at T-Systems. One hundred images per second are created in this way: “Because the volume of data is enormous at 10 gigabytes per second, we process all the data directly at the machine in an edge device. This reduces latency, and quality control takes place in real time.” The advantage of the innovation: The quality control takes place directly during the welding process rather than when the parts roll off the line. Instead of using spot checks to identify defective parts, all the welds that occur are inspected. Constantly, 100 percent of the time. This curbs the number of expensive and reputation-damaging recall actions, and the automated process speeds up production. The cost advantage is enormous: AI cuts quality control costs by up to 90 percent.
These types of applications are an important step on the way to a fully-automated and constantly self-improving factory. The joint research results from OEM and T-Systems were so clear that they have now flowed into concrete T-Systems solutions: The quality control of weld seams using so-called computer vision image processing is one of the core applications of the AI Solution Factory. T-Systems presented it for the first time at Hannover Messe 2021. The first automakers are already using it. But with the support of the AI Solution Factory, companies can just as easily optimize their logistics, for example, to better utilize their warehouses, cargo holds, and containers. Security monitoring in companies and in public spaces is another focus of the Factory: Companies can track vehicles based on computer vision services, for example. Or they can be informed automatically when unauthorized persons are in security areas or the AI discovers a helpless person. All personal data is immediately deleted from the edge device.
Regardless of the specific area of application, the AI Solution Factory provides its customers with a modular kit of hardware and software, managed cloud services, AI applications, cameras, and edge devices. The end-to-end service includes the development and testing as well as the implementation and operation of the AI solution. Cloud, connectivity, and security are included. This makes it easier for companies to deploy AI. Moreover, thanks to the AI Solution Factory framework, a completely new development is not necessary for every application. As a cross-cutting technology, AI promises new products and services, more efficient processes, and better decisions. For example, AI-based quality checks in real time, predictive maintenance, inventory tracking, and autonomous logistics can increase labor productivity by up to 40% in developed countries, according to studies. And specifically in manufacturing, AI enables greater plant efficiency, better machine availability, and even As-a-Service business models.
Word of the effectiveness of AI has spread throughout companies, according to Jörg Heizmann. More than 60 percent of CEOs worldwide now consider artificial intelligence to be more important than the Internet, but only 12 percent of companies have managed to productively implement AI systems in recent years, he said. “Everyone is talking about AI, but very few are taking advantage of the opportunities it offers.” Despite a positive proof of concept, 90 percent of AI projects are currently falling by the wayside for a variety of reasons.
If AI is a promise, why do companies so rarely deliver on it? Because the devil is in the design, or in data, security, and privacy. Take data, for example: Many companies collect data, but not always the right data. For example, to take quality control to a new level with machine learning, “it's not enough for companies to have noticed an error on day X,” Heizmann says. The data would also need to tell them what type of defect it was and what the engineering teams did to fix it”. This kind of information is rarely available; what is also missing in many places is the important sensor data.” Where the data stock leaves something to be desired, T-Systems first gathers the necessary information and takes care of the appropriate interfaces at the customer's site before developing a proof of concept together with the customer and taking care of the modeling and the training of the algorithms.
Without competent support, companies often drive their AI projects into the wall, even if they have a solid data set. “Many cannot meet data protection and data security requirements or they lack a coherent rollout strategy,” Heizmann says. “The biggest danger with artificial intelligence is that people think they understand AI far too early,” says AI researcher Eliezer Yudkowsky, summing up the difficulties. In terms of companies, this means that they rush ahead with AI projects without having the necessary preconditions. As a result, many fail to integrate artificial intelligence into their business processes, IT systems, and data sources. Or they rely on a solution that is neither scalable nor highly available.
The AI Solution Factory wants to prevent such failures and therefore offers companies a kind of turnkey AI. Behind it is a team of experienced AI and technology experts. The AI Solution Factory combines a deep learning factory, which provides data scientists with a standardized environment in the cloud, with an application suite, which makes connections to business systems and logics. “Our solutions are compatible with clouds from all vendors,” Heizmann says. “The machine learning tools are based on open-source technology, so we avoid lock-in effects.”
At the ITS World Congress 2021 in Hamburg, the AI Solution Factory demonstrated how it uses AI to increase safety in public transport while also taking into account data protection. Cameras already provide images in many buses and trains today. With the AI Solution Factory, T-Systems can analyze these “on the edge” using machine learning techniques. For a transport company in the Rhineland, the AI evaluates the passenger volume and seat occupancy and automatically recognizes dangerous situations. But don't transport operators risk creating data protection problems by doing this? “No” reassures Heizmann, "because thanks to the edge, the images can be analyzed in real time on site and don't have to go to the cloud. The camera images of public transport users are not stored.” With the T-Systems solution, passengers' privacy rights are protected.
