FAMILIAR

The term of federated machine learning (FedML) is popular in the context of public funded R&D projects. Still, it is rarely used in the industry especially not combined with other leading technologies such as XR and AM. But why? Currently, 2 approaches are considered. 1) FedML with homomorphic encrypted data – uses to much computing power, energy and produces large amount of traffic. 2) FedML by sending parameters of models to a controller – the controller must be a trusted third party, because he can reverse engineer the average data set of each client. Both approaches come with significant disadvantages. Further challenges include synchronous high-quality data acquisition, process-conform interpretation, and delivery of training parameters to the FedML model. FAMILIAR will be based on the second approach. The goal is to eliminate the controller (trusted third party) with the help of the gossip approach. In the created mesh it has to be possible to decide who is entitle to participate. Accordingly, the identity of the edge device must be known. This will be ensured by a blockchain-based approach. The solution shall be embedded and tested in real-life applications, such as Railway Engineering, Maintenance and Training, Welding and Human Robot Collaboration.To establish the use cases, sophisticated data mining techniques will be combined with deep learning. In conjunction with XR, it shall be possible to acquire and interpret high quality data, while providing the ideal parameters to the FedML model.