Remaining Useful Life Prediction of Rolling Bearings Based on Global-Local Attention Mechanism Transformer with Multi-Source Feature-Weighted Kernel Principal Component

Aiming at the health index (HI) is difficult to characterize when the rolling bearing state changes and the transformer network attention mechanism only focuses on the global feature information, resulting in the low prediction performance. This paper proposes a bearing RUL prediction method based on the global-local attention mechanism transformer with a multi-source feature-weighted kernel principal component. Squeeze and excitation networks (SENet) are applied to the adaptive assign weights to its multi-channel features. It combined with the distances of the multi-source features, which were calculated by the Mahalanobis distance (MD), to obtain the multi-source HI through the kernel principal component analysis algorithm (KPCA). It can balance multi-source features and retain important feature information. The global-local attention mechanism transformer method fuse and learn global and local spatiotemporal sequence features can improve bearing RUL prediction model performance. The proposed method was verified by experiments of the XJTU-SY dataset and the PHM2012 dataset and compared with the publicized methods. Results demonstrate the effectiveness and superiority of the proposed approach in terms of prediction accuracy.