As a key equipment in coal mine underground transportation systems, the bearings of mining scraper conveyors are prone to failure under extreme operating conditions such as high temperature, high impact load, dust pollution, and long-term operation. However, under complex operating conditions and multiple failure modes, existing fault diagnosis methods still face many challenges in terms of feature extraction and classification accuracy. To this end, this paper proposes a bearing fault diagnosis method based on a dual-branch neural network. Firstly, the one-dimensional vibration signal is converted into a two-dimensional time-frequency map using continuous wavelet transform to enrich the time-frequency features of the input data. Subsequently, a dual-branch neural network architecture is constructed, where one branch is based on a dilated convolution residual module to enhance local feature extraction capability, and the other branch combines a global attention mechanism to optimize the learning of global features. Finally, the diagnostic ability of the model is improved through feature fusion. Experimental verification is conducted on the CWRU bearing dataset and the PU bearing dataset, and the results show that the accuracy of this method in fault diagnosis reaches 99.78% and 97.94%, respectively. The overall performance is superior to that of comparative models such as DCRB-CNN, GAM-CNN, and GAM-ResNet, demonstrating significant diagnostic capability and good generalization effect. In addition, this paper also conducts bearing fault diagnosis experiments using coal mine field data, further verifying the practicality and engineering deployment value of the proposed method under strong noise and non-stationary operating conditions. This research provides an efficient and reliable deep learning solution for intelligent fault diagnosis of complex coal mine equipment such as scraper conveyors.