Intelligent fault diagnosis method for rolling bearings based on flexible residual neural network

Rolling bearings play a crucial role in rotating machinery, and their efficient operation is vital for the machine’s longevity and performance. In numerous real-world situations, diagnosing faults in rolling bearings presents significant challenges. Signals obtained from industrial applications often contain unavoidable noise, complicating analysis. Additionally, the intricate working conditions in actual operations can greatly influence bearing signal characteristics. Consequently, traditional diagnostic techniques struggle to effectively handle the effects of varying loads and noise. To improve the accuracy of fault diagnosis for rolling bearings in noisy and variable working conditions, a new approach using a flexible residual neural network (FResNet) is introduced. This network is built on a dynamic subtraction average-based optimizer (DSABO) and a parallel attention module (PAM). The core of FResNet is a flexible residual module based on convolutional neural networks, which allows for adjustments in the number of convolutional layers, convolutional kernels, and skip connections during optimization. These design features improve the network’s ability to extract fault features and prevent degradation. Second, a DSABO with a dynamic position update strategy is proposed for parameter optimization of the above FResNet with the flexible residual module. This optimizer helps the model avoid being trapped in local optima, strengthening the fault diagnosis performance of the network. Third, a PAM is integrated, featuring convolutional layers that combine channel and spatial attention. This integration enhances data feature extraction by aligning it with rolling bearing operation data. Together, DSABO, PAM, and FResNet create an effective rolling bearing fault diagnosis model known as DSABO–PAM–FResNet. Finally, the feasibility and effectiveness of the proposed DSABO–PAM–FResNet model are validated using the rolling bearing fault dataset from Case Western Reserve University in the United States. The ablation experiments reveal that the DSABO model consistently achieves accuracies above 97% across different noise environments. This performance surpasses that of models using grey wolf optimizer (GWO), butterfly optimization algorithm (BOA), and whale optimization algorithm (WOA), indicating the excellent search capabilities of the DSABO proposed in this paper. In noisy environments, the model incorporating the PAM module consistently achieves fault recognition accuracies above 97%. This performance exceeds that of models using the efficient channel attention module (ECAM) and spatial attention module (SAM), demonstrating PAM’s excellent capability to highlight fault signals. In challenging environments with a signal-to-noise ratio of –6 dB, the proposed model achieves a fault diagnosis accuracy of 97.18%, proving its strong noise resistance. Under different load conditions of 0.75 kW, 1.5 kW, and 2.25 kW the proposed model maintains an average accuracy of 98.2% in environments with a −4 dB signal-to-noise ratio. This demonstrates the model’s excellent adaptability to variable working conditions. Comparison results demonstrated that DSABO-PAM-FResNet outperforms other intelligent fault diagnosis methods in terms of diagnostic accuracy, providing a new and effective intelligent method for rolling bearing fault diagnosis.