An Adaptive FMD-ConvTransNet Method for Weak Fault Feature Enhancement and Identification of Rolling Bearings under Strong-Noise and Non-Stationary Conditions

To address the problems that weak impulsive components of rolling bearing signals are easily overwhelmed by noise and that cross-scale feature representation remains insufficient under strong-noise and non-stationary operating conditions, a GFMD-ConvTransNet fault diagnosis framework is developed. First, an adaptive FMD parameter optimization mechanism based on an improved grey wolf optimizer is constructed to jointly optimize the number of modes, filter length, and total number of filters. By integrating hierarchical uniform initialization, a Logistic-gated cosine convergence factor, and a stagnation-triggered mutation refinement strategy, the proposed optimizer improves population diversity and global search capability, thereby enhancing the convergence stability of parameter optimization. Second, a squared-envelope-spectrum-based periodic impulse composite evaluation index (SEPCI) is defined as the objective function for FMD parameter optimization. Combined with KSES-based sensitive mode selection and weighted reconstruction, this strategy enhances periodic impulsive features and suppresses redundant noise. Finally, a ConvTransNet deep diagnostic model integrating multi-scale wide-kernel convolution and Transformer is constructed. This model captures both local fine-grained features and global temporal dependencies, enabling accurate intelligent identification of fault states. The proposed method achieves an average identification accuracy of 99.66% on the CWRU dataset, which is 2.36% and 6.74% higher than those of CNN and Transformer, respectively. It also maintains high diagnostic accuracy on the bearing fault simulation test rig under different noise levels. The average identification accuracies reach 100% and 99.91% under SNRs of −5 dB and 5 dB, respectively. These results demonstrate that the proposed method has good noise robustness and generalization capability under strong-noise and non-stationary operating conditions. It can effectively enhance weak fault feature representation and achieve high-accuracy robust diagnosis of rolling bearing faults.