Rolling bearing remaining useful life prediction via parameter-optimized VMD and hybrid TCN-GRU-self-attention network

To accurately predict the remaining useful life (RUL) of rolling bearings under strong noise interference, this paper proposes an RUL prediction method based on adaptive Variational Mode Decomposition (VMD) and a hybrid Temporal Convolutional Network-Gated Recurrent Unit-Self-Attention (TCN-GRU-SA) framework. First, a parameter-optimized VMD algorithm is developed by integrating the Grey Wolf Optimizer (GWO) with VMD to extract effective intrinsic mode components (IMFs) and reconstruct denoised signals, thereby mitigating the impact of strong background noise. Subsequently, time-domain degradation features are extracted from the reconstructed signals to generate more representative feature datasets. These degradation features are then fed into a parallel TCN-GRU-SA prediction model. To enhance the model’s generalization capability and RUL prediction performance, the proposed hybrid architecture combines a Temporal Convolutional Network (TCN), which captures local temporal patterns, a Gated Recurrent Unit (GRU) for modeling long-term dependencies, and a Self-Attention (SA) mechanism to prioritize critical degradation-related features. Experimental validation on the PHM2012 rolling bearing accelerated lifetime dataset demonstrates that the proposed method achieves superior noise robustness and prediction accuracy compared to existing approaches. Specifically, it reduces the root mean square error (RMSE) by 18.7 % and improves the coefficient of determination (R2) by 12.3 % under high-noise conditions, confirming its effectiveness in industrial predictive maintenance applications.