Improving the production of welded parts with artificial intelligence
When it comes to welded parts in submarine construction or shipbuilding at thyssenkrupp Marine Systems we lay high value on quality and safety. To ensure that our welded parts can withstand high pressure and corrosion we use artificial intelligence to detect weak spots and errors in advance to provide safety for crews of our customers worldwide. Jonas, working in AI operations, knows the advantages of artificial intelligence and data analysis to produce welded parts in ship and submarine building.
Everyday work life in artificial intelligence operations
For Jonas a typical day-to-day work does not exist, it looks different every day. “We organize our work inside the AI team in an agile matter”, Jonas emphasizes. He explains that a typical workday for him in AI development is like the day of a developer. Just like a developer does not only focus on coding Jonas also spends a lot of his time on data analysis. “You could say that 80% of an AI project is data and the rest is the actual AI modelling”, he explains.
Further development of AI in welding
Ready for Jonas’ deep dive into the projects he works on with his team?
When it comes to databased welding optimization or further development of AI and welding the core of this predictive project is to forecast the outcome of a welding bead. A bead is one layer of weld, and each bead has a different role. So many beads have different configurations within an entire welded component.
The artificial intelligence model currently takes the data from the welding machines at thyssenkrupp Marine Systems such as voltage current and wire speed and predicts if the current state of the welded part is prone to breakage or unsafe in any way. “If the state is labeled as unsafe by the AI, we can say with an accuracy of 90 to 95% that there is going to be an error or there is an error inside the beat”, the expert explains.
Real time optimization thanks to artificial intelligence
The extraordinary thing is that the prediction is directly sent to the helmet of the worker who is currently signed to the welding machine. This way real time optimization can be incorporated in the production of welded parts and increase the quality of our products. “Our current results from the tests in the manufacturing environment represent the theoretical background. We have seen that our workers produce fewer errors with the activated AI inside the helmet”, Jonas tells us.
What does AI look like for the worker at the welding machine? “The end user only sees a small box with a red light inside the helmet. Behind the scenes there are WIFI networks, MQTT servers and AI services. But for the worker we try to make it as easy as possible”, Jonas explains.
How does AI learn?
When it comes to the project Jonas and his team are working on one must specifically talk about supervised learning in context with the AI they have developed. Supervised learning involves input data and target data for example. The expert explains that the input data is always of the same sort within training and later in production as well. The target data is only used within the training phase.
The initial training of the AI look like this: input data is provided, and the AI model is assigned to categorize the data in two classes. Such a class can be for example the category true or the category false. “We compare this class with the target data we have acquired for the input data beforehand. If the AI is right, fine. If the AI is wrong, a method called back propagation comes into play. Since we know the right class, we can try to change the weights inside the model in favor of the right class”, Jonas explains. By the repetition of this process and this might mean million times, the AI system learns patterns for different classes and categories and can separate different data into mentioned classes and categories during production as well. By supplying the AI with enough data and repeating the process you teach the artificial intelligence to know when something looks right or wrong inside a welded part since all categories of errors have their own pattern. Therefore, pattern recognition is an integral part of artificial intelligence when it comes to optimizing the quality of welded parts in our production.
Cooperation with the University of Applied Sciences of Kiel
Enhancing the quality of welded parts using artificial intelligence and data analysis has been brought to life by a corporation with the University of Applied Science of Kiel with our colleagues at thyssenkrupp Marine Systems. The two partners had split the project into AI development and building the feedback system. The development of the artificial intelligence itself was mainly done by Jonas and his team. Whereas the development of the feedback system happens over at the university. Together, they are driving the advancement of artificial intelligence in welding. One example for this is the visual AI that can identify errors through video data by using thermal cameras.
“In order to teach AI, the recognition of error patterns and turn it into a reliable tool for production it takes much repetition and our team also faced challenges”, Jonas explains. These challenges include low data quality. In this case the team must start over and tune the data collection process. Sometimes the data preparation is also unable to handle new data and Jonas and his team have a bias inside the data set which breaks the AI model. But through persistent and systematic approach the industrial engineering team overcomes each one of these challenges day by day and improves the AI which increases the welding parts.
In addition to the challenges, the support of artificial intelligence also offers advantages. The main benefit of using AI in the production of welded parts for ship and submarine building is that Jonas and his team can analyze not only the big but also the small changes within the welding process itself. This increases the quality of the welded parts, and the workers have less rework to do which in turn saves time money and resources. “On the other hand, we have less quality protocol work in our quality department since there are less detects that need to be protocoled.”, the expert adds.
What does the future hold?
In our fast-moving digital world artificial intelligence is something that is in a constant state of improvement. The more data points and repetition processes the AI tool of Jonas, and his team goes through the better and more accurate it will become. Therefore, the experts have developed an agile road map for the upcoming years in which they will be focusing on improving the manufacturing process with the artificial intelligence tool they have developed.
These optimizations will include the defect detection with visual AI tools as well as further predictive quality projects within additive manufacturing and fuel cell production. The expert adds: “My personal goals for the team are that everyone has fun building innovative AI products and that we can support and automate many manufacturing processes with our products to help the people in our company with repetitive, dangerous or exhausting tasks to help make this company a great place to work.”
The world of digitalization and artificial intelligence fascinates you and you want to become a part of our mission? Come on board and take a look at our current vacancies on our career page.