Objective In industrial production the operating environment of rolling bearings is complex and environmental noise reduces the accuracy of common network models for bearing fault identification leading to low noise interference resistance. To improve model identification accuracy and noise interference resistance a multi-branch sparse residual network was designed for bearing fault diagnosis. Methods The multi-branch sparse residual network utilized a multi-branch structure employing different convolution kernels to perform multi-scale feature fusion on features of varyingwindow sizes thereby enhancing the model?? s identification accuracy and resistance to noise interference. Additionally an attention mechanism was integrated into the multi-branch network to adjust the weights of the convolutional layers in each branch. Finally to improve the feature learning capability of the proposed method sparse residual connections were introduced to avoid data redundancy. Results Through comparative experiments the effectiveness of the proposed method was verified achieving an identification accuracy of 99. 2% under noise-free conditions and an identification accuracy of 95. 85% under strong noise conditions. In actual production scenarios the model?? s recognition rate was 98. 90%. Conclusion The experimental results indicate that the model has strong feature learning capability. It achieves high accuracy in bearing fault identification demonstrates strong resistance to noise interference and possesses significant practical value.