Data-driven nondestructive evaluation for bearing prognostics using a dual correlation graph knowledge distillation network

Bearings are critical components in mechanical systems, making vibration-based non-destructive testing (NDT) essential for continuous condition monitoring. Accurate remaining useful life (RUL) prediction based on these NDT data streams is crucial for ensuring operational reliability. However, existing deep learning models often encounter limitations in computational efficiency and reliability, hindering their practical deployment in NDT systems. To address these challenges, a dual correlation graph knowledge distillation network (DCGKDN) is proposed for RUL prediction. First, a dynamic graph structure utilising an adaptive adjacency matrix is constructed, allowing node correlations to be automatically learned from vibration NDT data. Next, a multi-level spatiotemporal knowledge distillation strategy transfers rich signal representations from a teacher model to a student model; consequently, model complexity is significantly reduced, while comparable accuracy is maintained. Finally, prediction reliability is quantified via a dual correlation uncertainty mechanism to account for signal noise and model variability. The effectiveness and superiority of the proposed method are demonstrated through experimental validation on two benchmark datasets.