Real-time update and prediction of local defect crack size in rolling bearings based on digital twin

Abstract

Addressing the challenge of real-time monitoring for local defect size in small-sample rolling bearings, this study proposes a digital twin-based method for dynamic updating and prediction of local defect crack size in rolling bearings. First, a digital twin model is built, integrating the Aquila optimizer and dynamics model to simulate real-time crack data. Second, a time-series Transformer model maps time-domain signals to crack data. A logistic regression model evaluates operational reliability. Additionally, an integrated Mixed Kernel Relevance Vector Machine and Bayesian Optimization Algorithm-Bidirectional Gated Recurrent Unit model predicts crack size. The method is validated using a bearing dataset, achieving low absolute and relative mean errors (0.1392 and 0.0102). It provides an effective solution for real-time crack monitoring and operational assessment.