Research on rolling bearing fault diagnosis based on parallel depthwise separable ResNet neural network with attention mechanism

An accurate and efficient fault diagnosis method for rolling bearings significantly contributes to enhancing the safety and stability of mechanical systems. To address the limitations of existing intelligent diagnostic methods—such as low accuracy and efficiency under complex operating conditions, excessive model parameters leading to prolonged inference times, a high risk of overfitting, and poor robustness—a parallel deep separable ResNet neural network based on an attention mechanism (PDSResNet-AM) is proposed. The model leverages a dual-input strategy by transforming one-dimensional vibration signals into two-dimensional time–frequency representations using both Continuous Wavelet Transform (CWT) and Short-Time Fourier Transform (STFT), thereby enriching the input features. Additionally, Depthwise Separable Convolution (DSConv) and the Convolutional Block Attention Module (CBAM) are utilized to enhance relevant features, reduce the number of parameters, and mitigate the risk of overfitting. Furthermore, an optimized dilated residual convolution module is introduced to replace conventional convolutional modules, enhancing the model’s generalization capability. Extensive experiments conducted under varying noise levels, load conditions, and rotational speeds demonstrate that the proposed method achieves superior diagnostic accuracy, strong generalization capability, and noise robustness compared to existing deep learning models. These results underscore the feasibility of deploying PDSResNet-AM in industrial applications.