Federated Transfer Learning with a Hybrid CNN-Transformer Model for Bearing Fault Diagnosis

Aeroengines operate under extreme conditions, making the development of robust fault diagnosis systems critical for ensuring operational safety and reliability. While deep learning models offer powerful feature extraction capabilities, their application is often hindered by data privacy regulations and proprietary concerns, which create isolated data silos and prevent centralized model training. To overcome these limitations, this paper presents a novel federated transfer learning framework designed for accurate and private cross-domain aeroengine diagnostics. The proposed framework deploys a hybrid CNN-Transformer model on each local client to effectively capture both local spatial patterns and long-range temporal dependencies from raw vibration signals. At the client level, a composite loss function incorporating maximum mean discrepancy is designed to proactively minimize the domain shift between different operational conditions. Subsequently, the federated averaging algorithm securely aggregates the learned model parameters from distributed clients, enabling collaborative knowledge sharing without exchanging sensitive raw data. Experimental results on two public bearing datasets confirm the efficacy of the proposed method, showing its clear superiority over existing baselines in terms of diagnostic accuracy.