Manipulating Feature Space for Generalizable Data Characterization in Federated Learning

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recently, federated learning (FL) emerged as a promising solution in the field of condition monitoring (CM) that allows collaborative model training across various machines without sharing raw data. However, variable working conditions often lead to discrepancies between clients, that leads to the formation of condition-dependent and randomly dispersed feature clusters on the global server. To address the above-mentioned issue, we have investigated various loss functions to enhance the organization and discriminative capability of the global feature space. First, a contrastive loss with cross-entropy is employed to improve inter-class separability under supervised learning. Second, a triplet loss with cross-entropy is applied to strengthen relative distance learning between positive and negative samples. Finally, clustering loss is in combination with center-pull optimization introduced to enable the server to form condition-invariant and compact feature clusters across different working states. For this purpose, a feature clustering technique is utilized. Our investigation shows that combining feature clustering with a centre loss function achieves promising results. The proposed approach enables the server to reorganize unsupervised features so that similar health states form distinct and compact feature clusters regardless of operating conditions. Three different scenarios are evaluated, including baseline condition separation, cross-condition exposure, and full-condition overlap, to assess the generalization capability of the proposed framework. In addition, two further validation cases are considered, incorporating more diverse fault categories and a larger number of clients. The proposed framework consistently strengthens inter-class boundaries, improves cluster compactness, and enhances classification stability across heterogeneous operating settings. These results demonstrate that the approach provides a scalable and reliable solution for cross-condition CM in federated environments.