A dual-channel multi-scale residual network with dilated convolution optimization for rolling bearing fault diagnosis

Convolutional neural network (CNN) as a core component of deep learning is utilized to effectively extract potential features from input data through multiple layers of nonlinear units and have thus been widely applied in bearing fault diagnosis. However, the number of model parameters is significantly increased as the depth of the network increases, which potentially leads to overfitting during training and results in negative impacts on the model’s performance. Furthermore, introducing spatial invariance in convolution and pooling operations may reduce the model’s sensitivity to small changes in input data, impairing its ability to capture detailed information. To address these issues, a dual-channel multi-scale residual network with dilated convolution optimization for rolling bearing fault diagnosis is proposed by fusing the Inception network, Residual Neural Network (ResNet), and Squeeze-and-Excitation Network (SENet) attention mechanism to improve CNN. The method integrates the optimized multi-scale feature extraction module in the Inception network and ResNet, which enhances the feature extraction capability of the model, improves computational efficiency, and reduces performance degradation. Additionally, an SENet channel attention mechanism is incorporated to emphasize important features while suppressing irrelevant information, overcoming the network’s limitations in capturing detailed information. Furthermore, a two-branch network structure with parallel inputs of original and noisy STFT time-frequency images is designed and combined with dilated convolution optimization. The structure is designed not only to effectively extract time–frequency features and mitigate noise interference but also to enhance the model’s generalization ability, reduce parameter complexity, and enhance diagnostic accuracy. The effectiveness of each module is validated through ablation experiments on a public dataset. It is demonstrated that the proposed method outperforms four other neural network models in terms of accuracy, establishing its suitability for bearing fault diagnosis tasks.