Mapping of Newcomer Clients in Federated Learning Based on Activation Strength

Federated learning is a collaborative machine learning approach that allows multiple parties to train a model without exchanging sensitive data. In manufacturing, where different parties may have proprietary or sensitive data that cannot be shared, this is especially useful. However, traditional federated learning approaches (as proposed by McMahan et al.) do not consider the differences in data and computing resources across different parties, leading to sub-optimal model performance. Personalized federated learning addresses this issue by allowing each party to contribute to the model training according to its specific data and resources. Furthermore, most common approaches only consider a limited set of data and a short period of time, without considering the system’s long-term usefulness. It is important to consider the integration of new clients and the continuous change of data, which could result in the addition of new classes. This paper will explore the potential of federated learning in manufacturing and present a flexible and expandable approach, focusing on mapping newcomer clients based on activation strength of weights.