Saving raw materials and energy in manufacturing processes is the motto of the day. This also applies to welding in the metal sector. It is important that welding errors (often user errors) are recognised quickly and automatically. Artificial intelligence (AI) can help with this. Fraunhofer IPA has developed an AI concept for the company Lorch.
Skilled technicians are scarce. Companies therefore often have to rely on inexperienced staff. This increases the likelihood of user errors, resulting in more faulty products being produced and thus more wasted material and energy. Artificial intelligence can recognise such operator errors and errors caused by wear processes at an early stage and thus reduce them. Welding errors can also be reduced. However, AI systems need a lot of data. Indeed, they must first be trained with the relevant data. This is where another problem arises. Companies that use systems from a manufacturer, for example, usually do not want to hand over this data just like that.
Preventing welding errors with artificial intelligence
Lorch Schweißtechnik also faced this problem. It therefore engaged Fraunhofer IPA. How, they asked, can user errors in welding processes be reliably recognised via AI without customers having to hand over their sensitive welding data? Fraunhofer IPA's answer: with federated learning, or collaborative learning. It is a machine learning technique in which an algorithm is trained through multiple independent sessions, each with its own dataset. "The special thing about this is that we train the artificial intelligence with the customer's data without the data leaving the company in question." So says Can Kaymakci, a scientist at Fraunhofer IPA. The trick is that each customer uses its data to train its own AI model. It is not the data that is exchanged, just the AI models. These are combined into a single, better-optimised overarching model.
Suitable AI model
First of all, Fraunhofer IPA researchers had to select a suitable AI model for detecting energy anomalies. They had to select a model that recognises user errors based mainly on energy consumption data. To this end, they collected data on the welding process to be observed in the Lorch laboratory. They collected the data including the intentional inclusion of "user errors".
200 welding tests
They carried out about 200 welding tests. A lot, but not enough to train an artificial intelligence. "That's why we multiplied the data, making the original 200 datasets 2,200," Kaymakci explains. Photos are the best example of how this works. You can rotate them, mirror them, convert them to black and white, zoom them and so on. That way, you can generate a lot more data.
User errors
The team also investigated how many measurements per second are needed to reliably recognise user errors. The result: fewer measurement points are enough than expected. "This way, we can reduce the storage capacity required, simplify communication and process less data, which in turn saves time, cost and energy," Kaymakci summarises. The researchers implemented the model they created on a Lorch welding power source.
Errors are quickly recognised
What are the benefits of federated or collaborative learning? The researchers used a specially developed simulation tool to answer this question. They analysed three scenarios. First, an artificial intelligence trained with all customer data. This is a hypothetical assumption because this data is not available to the welding machine manufacturer. Second, models that were trained only with a single customer's data. And third, federated learning. This involves merging the models of customers.
Results
“The results speak for themselves. The recognition rate of a model trained using collaborative learning is 0.81. This result is comparable to that of a system for which all customer data was available for training. By contrast, systems trained solely on data from a single customer achieve a recognition rate of just 0.45,” confirms Kaymakci. For welding equipment manufacturer Lorch, this means that in future it will be able to offer its customers added value via the AI system without the need to store the data centrally at Lorch. Customers, in turn, will benefit from being able to identify errors more quickly and draw on the ‘knowledge’ of all customers.
Not only welding processes
Of course, this form of artificial intelligence can be used for more than just recognising welding errors during welding processes. Rather, the system is suitable for any issue where artificial intelligence adds value but the data required to do so is sensitive.
Opening photo: Lorch's laboratory collects data from the welding process to train artificial intelligence to recognise welding errors automatically and quickly. (Photo: Lorch)
Read also: More sustainable powder coating with artificial intelligence
