The problem: Today, many machine builders, users or customers still don't want to share data with manufacturers and at the same time demand optimizations of their plants - also using machine learning. But without sufficient data, no training and no models, and thus no machine learning. Experts speak of the Distributed Data Dilemma - the user cannot get the data he needs to solve his problem.

The solution: U.S. tech companies have known about the problem of mechanical engineering, industry for a long time. The smartphone industry is considered a pioneer for the federated learning approach. If a smartphone customer presses the letter A in his message app, the words "but or so" appear. If they don't fit, he continues typing. If he presses A and L, the system suggests the word "all." He selects it and continues typing "all in" and the term "order" is immediately given by the application. And it is not only his smartphone that does this, but all Google or Apple devices around the world. They have been unknowingly trained locally by the user. If he now selects a suggested word, he continues to train the model on the smartphone. In the evening, when the phone is charged, the device trains the models and makes them ready for dispatch. As soon as the smartphone is back online, it sends the trained model to the Google or Apple developers. The user does not send any data, but encrypted trained models. All information entered on the smartphone does not leave the smartphone. The developers use this for their models and then send them back newly trained. A continuous improvement cycle is created without data being shared. And this principle also works for machines.

The challenge: In federated learning, the first step is not the AI models, but the infrastructure and development tools that make it possible to work with data to which there is no direct access. Distributed networks with edge devices, cloud connections, and then in different geographic locations with limited connectivity are everyday occurrences. In addition, providers like Katulu must first find all the assets necessary for learning in production and group them in a meaningful way, without directly learning anything about the assets.

The basic model: The next step is to create an initial model. For this, employees of the machine builder record initial data sets of the machine in the technical center or laboratory. The customer's models are also encrypted so that no conclusions can be drawn from the data.

The project partners explain how and why Siemens is using the technology and what experience they have had with it in the AI in Industry workshop.