Abstract: A rolling bearing is one of the most important components in the rotating machinery and equipment of agricultural production. Most rolling bearings can often operate under complex conditions, such as strong noise and variable working conditions. It is very necessary to fully extract the characteristics of the fault signal in the process of fault diagnosis of rolling bearings. However, the convolutional neural network cannot fully extract the global feature information during the feature extraction under variable working conditions of rolling bearings. In this study, a multiscale convolutional neural network (MSCNN) - swin transformer (SwinT) was proposed for the rolling bearing fault diagnosis. Firstly, the continuous wavelet transform (CWT) was used in the data processing module to convert the one-dimensional vibration signal into a two-dimensional time-frequency image, which retained the time-frequency characteristics of the original signal. Secondly, the MSCNN was used to extract the local features of fault information in the local perception module. Then, the key information was extracted by the convolutional block attention module (CBAM). As such, the feature extraction module was further constructed. The residual connection was introduced to improve the utilization efficiency of the feature information before and after extraction. The global features of fault information were learned using the SwinT network. Finally, the global average pooling was used instead of the fully connected layer for fault identification. The bearing dataset was collected from the Case Western Reserve University and the self-made experimental data for the experimental verification. The test results showed that the fault identification accuracy in the visualization, variable working condition, and the test of different coding tests were 99.67%, 95.01%-99.66%, and 100%, respectively. The accuracy of fault diagnosis reached 99.18% in the self-made experimental datasets. Compared with the CWT-LeNet5, CWT-VGG16, CWT-ResNet18 and CWT-SwinT, the average fault identification accuracy was improved by 8.79, 8.64, 3.49, and 3.18 percentage points, respectively in the variable working condition test, while by 5.23, 2.74, 1.40, and 1.26 percentage points in the self-made experimental data set, respectively. The accurate and rapid identification was realized in the different fault states of rolling bearings under complex conditions, such as variable working conditions. The global characteristic information of bearing faults was fully extracted with high fault diagnosis accuracy and better generalization ability. This finding can also provide a strong reference for the rolling bearing fault diagnosis under variable working conditions.