Federated learning makes data exchange unnecessary
Lorch Schweisstechnik GmbH was also faced with this problem – and therefore brought Fraunhofer IPA on board. How, is the question, can user errors in welding processes be reliably detected using AI without customers having to hand over their sensitive welding data?
Fraunhofer IPA’s answer: With the federated learning approach. “The special thing about it: We train the artificial intelligence with the customers’ data without the data leaving the respective company,” says Can Kaymakci, scientist at Fraunhofer IPA. The highlight is that each customer trains their own AI model with their data – it is not the data that is exchanged, but only the AI models. These are combined into a single, better optimized overall model.
First of all, the researchers at Fraunhofer IPA had to select a suitable AI model for energy anomaly detection – a model that detects user errors primarily through energy consumption data. To do this, they collected data in the Lorch laboratory about the welding process being observed, including the intentional inclusion of “user errors”. They carried out around 200 welding tests. A lot, but not enough to train an artificial intelligence. “We therefore duplicated the data; the original 200 data sets became 2,200,” explains Kaymakci.
How this works can best be understood using photos as an example: you can rotate them, mirror them, convert them to black and white, change the zoom – and in this way generate more data. The team also investigated how many measurements per second are necessary to reliably detect user errors.
The result: fewer measuring points are sufficient than expected. “In this way, we can reduce the required storage capacity, simplify communication and process less data, which in turn saves time, costs and energy,” summarizes Kaymakci. The researchers implemented the model they created on a welding power source from Lorch.