Filtered contrastive learning with kernel attention and its application in rolling bearing fault diagnosis

Contrastive learning (CL), as a commonly used method for mechanical fault diagnosis, has achieved good application results, but it is prone to introducing false negatives and redundant easy negatives, especially in strong noise environments. To address these limitations, a novel framework, named filtered CL with kernel attention (FCL-KA), is proposed in this paper. Specifically, a new spatially adaptive encoding module is constructed, which can enable dynamic receptive field calibration for precise feature capture under high-noise interference, thus overcoming the limitations of fixed receptive fields and capturing discriminative features from noisy signals. Furthermore, a dual-threshold filtered contrastive mechanism is designed to explicitly eliminate high-similarity false negatives, screen out low-information easy negatives, and build more reliable prediction models. Finally, experimental verification was conducted on two rolling bearing fault datasets, and the results showed that FCL-KA exhibited superior classification and diagnostic performance in different evaluation metrics.