Intelligent diagnosis of gearbox in data heterogeneous environments based on federated supervised contrastive learning framework

To address the model training bottleneck caused by the coupling of data silos and heterogeneity in intelligent fault diagnosis, this study proposes a Federated Supervised Contrastive Learning (FSCL) framework. Traditional methods face dual challenges: on one hand, the scarcity of fault samples in industrial scenarios and the privacy barriers to cross-institutional data sharing result in insufficient data for individual entities; on the other hand, the data heterogeneity caused by differences in equipment operating conditions significantly diminishes the model aggregation effectiveness in federated learning. To tackle these issues, FSCL integrates the federated learning paradigm with a supervised contrastive mechanism: firstly, it overcomes the limitations of data silos through distributed collaborative training, enabling multiple participants to jointly develop diagnostic models without disclosing raw data; secondly, to address the feature space mismatch induced by heterogeneous data, a hybrid contrastive loss function is designed, which constrains the similarity between local models and the global model through supervised loss, thereby enhancing the feature representation capability of the global model. Experiments on two gearbox datasets demonstrate that the FSCL framework effectively resolves the issues of data insufficiency and heterogeneity, providing a novel approach for intelligent maintenance of industrial equipment that optimizes both data efficiency and privacy protection.