Hybrid real-virtual deep CCD method for fault diagnosis of mining bearings

Abstract

In mining applications, rolling bearing fault diagnosis requires not only fault type classification but also reliable assessment of severity and defect depth. Conventional regression methods often fail under mining conditions, where signals are highly non-stationary, noise levels are severe, and labeled data are scarce. To address these challenges, a Hybrid Real-Virtual Deep CCD method (HRV-DeepCCD) is proposed. The framework integrates real and simulated data, combines wavelet packet decomposition and one-dimensional convolutional networks for feature extraction, and applies a Center-Constrained Distance loss to impose physically meaningful constraints. By introducing a classification-guided interpolation mechanism to replace direct regression, the method achieves stable depth estimation under adverse conditions. Experiments conducted on conveyor belt bearings in underground coal mines demonstrate a classification accuracy of 99.44 %, with regression errors reduced by 80.7 % (RMSE) and 73.0 % (MAE) compared with the LightGBM baseline. These results confirm the framework’s high accuracy and robustness, providing a practical and noise-resistant diagnostic solution for harsh mining environments.