Patch-Based Fourier Attention-Enhanced Contrastive Learning Networks for Robust Drift Diagnosis in Long-Sequence Bearing Data

In industrial applications, the non-stationary nature of long time series data from bearing operations poses a significant challenge due to data drift, influenced by varying operating conditions and environments. To tackle this issue, we propose a novel fault diagnosis model leveraging contrastive learning. This approach utilizes domain-differentiated contrastive feature learning to construct positive and negative sample pairs, fully capturing the commonalities within the same fault type and the differences between different fault types, thereby enhancing the model’s robustness against interference. Moreover, the model employs a patch-based transformer to capture dependencies within local subsequences, reducing computational complexity while maintaining the ability to abstract comprehensive signal representations. Additionally, the integration of multi-head fourier attention allows simultaneous analysis of time-domain and frequency-domain characteristics, enriching the feature extraction process. Our method is validated through comparative, parameter analysis, and ablation studies on datasets, demonstrating its effectiveness and potential for improving fault diagnosis accuracy in bearing systems, thereby reducing downtime and enhancing operational safety.