Aiming at the problem that the convolutional neural network extracts the features from the input signal through the local receptive field, and cannot effectively capture the global context information under variable load and noise environments, resulting in the low recognition accuracy of rolling bearing fault diagnosis, a rolling bearing fault diagnosis method based on multiscale adaptive depthwise separable convolution (MADSC) and spatial interaction double-stream Swin Transformer (SIDSwinT) is proposed. Firstly, one-dimensional vibration signals are converted into two-dimensional time-frequency maps using wavelet transform to retain the complete information. Next, MADSC is constructed to extract local feature information and capture the characteristic changes of rolling bearing vibration signals at different scales. After that, SIDSwinT is designed to extract the global feature information, and the proposed spatial interaction module (SIM) is utilized to adaptively adjust the feature weights, while the sampled information is weighted by the deformable attention to eliminate the distributional differences caused by fluctuations in working conditions. Finally, bidirectional long short-term memory (BiLSTM) is utilized to better understand the contextual information and to improve the diagnostic accuracy and stability. Two different datasets are used to verify the fault diagnosis performance of the proposed method, and the experimental results show that the accuracy of the proposed method is higher than 93.00% when the signal-to-noise ratio is -4, and the accuracy is higher than 92.00% under the condition of variable load, which verifies that the proposed method has a stronger anti-noise performance and generalization ability than the comparison methods.