Convolutional neural networks(CNNs) show significant potential for bearing fault diagnosis. However, tra-ditional CNNs face challenges such as poor noise resistance, high computational complexity, reliance on extensive samples, and limited generalizability. As a result, this paper proposes WDSC-Net, a lightweight, multiscale feature fusion method, focusing on limited labeled fault samples. Initially, a wide kernel convo-lutional is employed, aiming to reduce parameters and computational complexity. Next, features are fed into a 1×1 convolutional layer reduces feature dimensionality. Subsequently, leveraging the benefits of depth-separable convolution (DSC) allows the separation of spatial and channel features, constructing four convolutional layers of varying scales to amplify the nonlinear fault representation. Finally, an improved feature soft-threshold denoising module is introduced for global feature denoising. Validation on CWRU and MCDS datasets shows that the WDSC-Net method exhibits superior generalizability and noise resistance compared to typical deep-learning fault methods