Cross-domain federated learning with self-attention collaboration: efficient mechanical fault diagnosis for data island scenarios

Deep learning enables effective autonomous fault feature extraction, but the increasing shift toward decentralized data storage, driven by stricter privacy regulations, has led to data silos that hinder its full potential. Federated learning (FL) offers a promising solution; however, existing FL methods typically rely on conventional networks and fail to capture the complex, dynamic nature of federated data, resulting in suboptimal diagnostic performance. To address these challenges, this paper proposes a novel federated fault diagnosis model based on a self-attention mechanism. A self-attention network (SAN) is employed as the local model for each client, leveraging its selective memory and attention allocation capabilities to directly capture key global features, while avoiding the need for complex auxiliary structures. A global SAN is constructed by aggregating local models using the classical federated averaging (FedAvg) algorithm, which mitigates the adverse effects of data heterogeneity while preserving data privacy. The proposed method is thoroughly evaluated using a CNC spindle-bearing dataset under cross-speed, cross-load, and combined conditions. Comparative experiments with three widely used deep learning models in federated settings demonstrate that the proposed SAN-based federated model significantly improves diagnostic accuracy and generalization capability, effectively addressing the limitations of conventional FL approaches.