Physics-driven digital twin model and CNN-GRU network for remaining useful life prediction in rolling bearings

Abstract In order to enhance the accuracy and adaptability of remaining useful life (RUL) prediction for rolling bearings, this paper proposes a hybrid prediction framework that integrates digital twin (DT) modeling with a CNN-GRU deep learning architecture. First, the discrete wavelet transform is applied to denoise the measured vibration signals, and 13 multi-dimensional health indicators are extracted to characterize the bearing degradation features, whose relevance is further assessed using feature sensitivity and redundancy analysis. Subsequently, a bearing dynamic model and defect evolution model are established based on Hertzian contact theory, and the Snow Goose Algorithm is introduced for dynamic model updating, enabling lifecycle estimation of defect size and generating full-lifecycle simulated vibration signals. In the prediction stage, a CNN-GRU model is constructed, where the CNN convolutional layers extract local spatial features and the GRU units capture long-term dependencies in the time series, thereby enabling deep modeling of the bearing degradation process and RUL prediction. Experiments are conducted on the XJTU-SY bearing accelerated life dataset under three different data scenarios: ‘digital twin data only’, ‘hybrid data’, and ‘measured data only’. The results show that the proposed CNN-GRU model outperforms conventional CNN and GRU models across all error metrics. It effectively addresses the limitations of traditional methods under conditions of limited full-lifecycle bearing data. It provides reliable technical support for the intelligent maintenance and optimization of rolling bearings. The effectiveness and engineering potential of the DT-assisted RUL prediction method are thoroughly validated.