Condition-Adaptive Dynamic Evolution Method for Digital Twin of Rolling Bearings Based on Continual Learning

As a common and key part of rotating machinery, rolling bearings mainly determine the health of the equipment. Digital Twin (DT) can help comprehensively monitor and predict the remaining useful life (RUL) of rolling bearings by generating a high-fidelity model of the physical object. Due to the varying working conditions faced in real industrial applications, the existing DT model of rolling bearings urgently needs continuous dynamic evolution ability and resource-efficient updating method. In this work, we propose an end-edge-cloud based DT architecture for rolling bearings, which includes two stages: DT construction and DT update. For adapting to dynamic changes of physical entities under various working conditions, a Condition-Adaptive Dynamic Continual Learning Digital Twin Model (CADCL-DTM) is designed, which applies different penalties to continual learning tasks in both intra-condition and intercondition scenarios. Furthermore, we apply the proposed DT methods to RUL prediction of rolling bearings, enabling the DT model to continuously learn new conditions while mitigating catastrophic forgetting of old ones. Comprehensive experiments have been carried out with FEMTO-ST bearing to verify the dynamic updating ability and forgetting performance under the same working conditions and different working conditions, respectively.