TFFT-Net: A Hybrid CNN-Transformer Approach for Time-Frequency Feature Fusion in Bearing Fault Diagnosis

The industrial field environment is subject to dynamic fluctuations, yet deep learning models are typically trained on historical data, which may reduce their effectiveness in new contexts. This research introduces a Transformer-based temporal correlation embedding transfer learning approach to address the widespread issue of domain shift in bearing fault diagnosis. Specifically, the study presents a temporal correlation extraction module based on Transformer architecture, designed to capture sequential dependencies within vibration signals. Additionally, a multi-objective decomposition classifier is proposed to integrate and classify features across various spatial domains. The proposed methodology has shown superior performance in bearing fault diagnosis tasks, thereby demonstrating the algorithm’s efficacy and superiority.