To address the problem of poor bearing fault diagnosis due to the large amount of deep network feature parameters and the unbalanced number of fault category samples, this paper proposes a lightweight bearing fault diagnosis method under unbalanced data. Firstly, the one-dimensional vibration signals collected by the sensors are reconstructed into a two-dimensional grey scale map as model input.Secondly, an asymmetric multi-scale feature extraction module is designed to extract features from the input signal using convolution and null convolution at different scales, and a part of the features are mapped to the original space for removing noise and restoring the original data structure. Next, the extracted rich feature information is fed to the channel position bi-weighting module to bi-directionally weight the key channel and key position features using inverse channel convolution and local averaging. Then, a depthwise separable convolution (DSC) dense residual structure is designed to increase the feature fusion of each layer of the network while keeping the network lightweight and optimize the backpropagation performance through shortcut paths. Finally, the focal loss function is used to adjust the learning process of the model according to the importance of different fault categories, thus better adapting to the unbalanced data distribution.