To enhance the intelligence of locomotive sanding state recognition and address the limitations of traditional methods in terms of efficiency and cost, this paper proposes an image classification model that integrates multi-scale window dilated attention with an agent-enhanced mechanism. The model combines multi-scale window partitioning with dilated attention to improve the perception of locally sparse sand particle regions. A set of agent tokens is introduced to construct global semantic interaction paths, and a feature enhancement module is incorporated to increase feature diversity and discriminative capability. Experimental results on a self-constructed sanding image dataset demonstrate that the proposed model outperforms mainstream methods in terms of accuracy, recall, and F1 score, showing superior recognition performance and robustness. Furthermore, Grad-CAM++ visualization confirms the model's strong ability in localizing sanding regions and enhances its interpretability. The proposed approach offers a practical reference for intelligent maintenance of railway locomotives.