Digital twin-driven unsupervised waveform segmentation for bearing quantitative diagnosis

The quantitative diagnosis of bearing is a prerequisite for informed maintenance decisions, ensuring the high-efficiency operation of modern production facilities. Existing studies utilize dual-impulse extraction-based signal processing techniques or neural network-based intelligent methods for defect size estimation. However, the former is subject to expert knowledge and complicated interferences, while the latter is limited by data resources and black-box attributes. Simulation-based digital twin (DT) technology provides intrinsic mechanism insights and cost-effective data generation. Inspired by this, a DT-driven unsupervised waveform segmentation (DTUWS) method is proposed in this paper to address the above problems. Specifically, a high-fidelity DT model of bearing is first established based on the modeling-update concept of DT technology. The hyper-real observation capability of the DT model is leveraged to generate vibration responses and pixel-level fault semantic labels. Then, the U-Net structure is combined with multi-task learning to construct an unsupervised waveform segmentation model for feature extraction and knowledge transfer. The predicted semantic labels of unlabeled raw field signals are post-processed to derive defect sizes. The diagnosis mechanism of DTUWS is intuitive and interpretable. Experiments on two distinct bench tests demonstrate that DTUWS can achieve accurate and robust quantitative diagnosis without field pre-testing and manual feature extraction.