Abstract: To address the problems of low feature extraction accuracy, difficulty in balancing lightweight design and diagnostic accuracy, and insufficient topology information mining in rigid tank guide fault diagnosis, this paper proposed a diagnostic model combining dual-dimensional attention and a light pruning strategy. First, a dual-dimensional channel–coordinate attention mechanism was constructed to collaboratively analyze fault feature weights and spatial correlation attributes in both channel and coordinate dimensions, thereby improving feature extraction accuracy and suppressing noise interference. Second, a light pruning strategy was designed, which combined sparse channel screening based on lasso regression with real-time feature enhancement via dual-dimensional attention, achieving complementarity between sparse screening and real-time compensation while retaining key fault features and compressing the model scale. Finally, a topology enhancement module with pruning consistency was designed to enhance topology information mining and improve diagnostic accuracy through global normalization of feature dimensions, adaptive calibration of learnable affine parameters, and construction of sequential links. Experimental results showed that the average diagnostic accuracy of the proposed model after pruning reached 99.86%, and the memory consumption was reduced by 21.08% compared with the best-performing conventional model. Ablation experiments verified the synergistic gain of each module, and anti-noise tests demonstrated that the model had excellent anti-interference ability and could adapt to complex mine working conditions. This study can provide effective technical support for the accurate identification and lightweight deployment of rigid tank guide faults in smart mine and intelligent coal mine construction.