Bearing remaining useful life prediction based on periodic identification and expert knowledge embedding

Abstract

To improve the prediction accuracy of the remaining useful life of bearings, this study proposes a neural network model that incorporates expert knowledge embedding and periodic identification methods. In the proposed architecture, expert knowledge features are first extracted from raw vibration data and embedded into a high-dimensional space. A trend decomposition module is then implemented for data augmentation, followed by a periodic identification module to capture characteristic periodicities. Subsequently, a sequence-to-sequence architecture is adopted for the final remaining useful life prediction. The effectiveness of the model is validated through accelerated bearing life experiments. Experimental results demonstrate that the proposed approach achieves superior performance on the publicly available PHM2012 dataset, outperforming existing state-of-the-art methods. Furthermore, it obtains even better performance on a proprietary accelerated bearing life test dataset, confirming its strong generalization capability across varying operating conditions. Ablation studies further verify the individual contributions of the integrated modules to the overall robustness of the model. These findings establish an effective framework for predicting the remaining useful life of critical rotating components in mechanical systems, highlighting its significant potential for practical predictive maintenance applications.