Noise-robust fault diagnosis of elastic ring-supported rolling bearings using multi-scale dilated convolution and multi-head self-attention

Elastic ring-supported rolling bearings are critical components in high-end rotating machinery such as aero-engines and gas turbines. However, the vibration responses of such bearings are often affected by complex transmission paths, elastic-support-induced dynamic coupling, and strong background noise, which makes weak fault feature extraction and reliable diagnosis difficult. To address these challenges, this paper proposes a noise-robust fault diagnosis method termed the multi-scale dilated convolution and multi-head self-attention convolutional neural network (MSDA-CNN). In the proposed framework, Z-score normalization and variational mode decomposition (VMD) are used to preprocess raw vibration signals and construct multi-channel inputs that retain fault-related components in different frequency bands. A multi-scale dilated convolution module with parallel branches is then employed to extract complementary local and long-range temporal features under noisy conditions. Finally, a multi-head self-attention (MHSA) layer is introduced to enhance global feature interaction and improve class-discriminative representation learning. Experiments on a self-built rotor-bearing fault diagnosis platform and the public Paderborn University bearing dataset demonstrate that the proposed method achieves average diagnostic accuracies of 99.86% and 98.52%, respectively. Under different Gaussian noise levels, MSDA-CNN consistently outperforms MLP, 1D-CNN, WD-CNN, and MCL-DRL.