Intelligent fault diagnosis of bearings using dynamic weighted Markov transition field and hybrid attention ResNet

To address the insufficient characterization of long-range temporal dependencies in rolling bearing vibration signals and the indiscriminate treatment of impulse features and stationary components in conventional Markov Transition Field (MTF) encoding, this paper proposes a fault diagnosis method incorporating dual attention enhancement in both the signal and image domains. In the signal encoding stage, a Dynamic Weighted Markov Transition Field (DWMTF) is designed. This method first extracts the envelope of the raw signal via the Hilbert transform and then uses the envelope amplitude as a dynamic weight to perform amplitude modulation on the original signal. This operation adaptively amplifies the waveform characteristics at fault impact moments, thereby endowing the subsequently generated two-dimensional texture images with enhanced inter-class discriminability. In the image classification stage, a Hybrid Attention Residual Network (HAResNet) is constructed. The network introduces parallel channel attention (CA) and spatial attention (SA) branches into the ResNet backbone and adaptively fuses them through a nonlinear gating mechanism, thereby sharpening the network’s sensitivity to diagnostically relevant feature maps. Experiments conducted on the Case Western Reserve University (CWRU) bearing dataset demonstrate that the proposed DWMTF encoding achieves an average best classification accuracy of 98.61%, which is 3.95 percentage points higher than that of standard MTF. The proposed hybrid attention module outperforms single channel or spatial attention by 1.78 and 1.16 percentage points, respectively. t-SNE visualization and confusion matrix further verify the separability and reliability of the features extracted by the model. Comparisons with recent representative models including Transformer, DRSN, AHA-RTN, and MRAFDN suggest the favorable performance of HAResNet under the tested conditions. Robustness is validated under Laplacian noise at multiple signal-to-noise ratios. Cross-condition experiments across 1750, 1772, and 1797 rpm and cross-dataset validation on XJTU-SY demonstrate generalization and transferability. Hyperparameter sensitivity analysis for λ and Q, Grad-CAM visualisation, and computational efficiency evaluation further support the practical viability of the approach. This method provides an effective intelligent diagnosis solution for industrial bearing condition monitoring.