TL-SBRN: a Siamese BiGRU gated residual shrinkage transfer network for robust cross-domain rolling bearing fault diagnosis

Authors:

Abstract
Under noisy cross-domain conditions, limited temporal modeling and sensitivity to interference can reduce the reliability of learned feature representations and degrade diagnostic performance in bearing fault diagnosis, especially when only a small amount of labelled target-domain data is available. To address this issue, a Siamese bidirectional GRU (BiGRU) Gated Residual Shrinkage Transfer framework (TL-SBRN) is developed for robust cross-domain bearing fault diagnosis. TL-SBRN integrates temporal modeling and feature denoising within a Siamese learning structure, allowing these components to complement each other and improve feature robustness under noisy interference. Specifically, a shared BiGRU backbone is employed to capture forward and backward temporal dependencies from vibration sequences, while residual shrinkage blocks are embedded within the Siamese branches to suppress redundant noise-sensitive components before feature comparison. Source-domain pretraining followed by target-domain fine-tuning is further adopted to improve adaptability under limited labelled target-domain conditions. Experiments on laboratory-collected bearing data and electric locomotive bearing data demonstrate that TL-SBRN achieves robust and stable diagnostic performance under noisy cross-domain conditions. Supplementary few-shot evaluations further examine its applicability in severely data-constrained scenarios.

This publication has 30 references indexed in Scilit: