To address the insufficient feature extraction of existing fault diagnosis models when dealing with nonlinear, non-smooth signals in noisy underground environments—which greatly reduces diagnostic accuracy—this paper proposes a shearer rocker gear fault diagnosis model based on the dual-branch fusion of SwinT and SKNet, enabling comprehensive extraction of both global and local features. The model consists of two branches, the simplified SwinT and the lightweight SKNet. The SwinT branch is integrated with a BAM module to further strengthen its global feature extraction capability. In the SKNet branch, ordinary convolution is replaced with depthwise separable convolution, and its fully connected structure is substituted with one-dimensional convolution, which reduces model complexity, improves efficiency, and facilitates deployment on mobile and edge devices. The experimental validation is carried out by using the shearer rocker arm loading test bench of Tai Heavy Coal Machine, and the results show that the accuracy and precision of the proposed method are higher than that of the comparison model, and the recognition accuracy still maintains more than 97% when the signal-to-noise ratio is SNR=-6, which fully proves that the method in this paper has a strong noise immunity.