MCL: Mamba-Based Contrastive Learning method for coal gangue identification

Accurate identification of coal and gangue during top coal caving is essential for advancing intelligent and unmanned mining. Vibration signal analysis has shown strong potential in distinguishing coal from gangue; however, effectively capturing subtle differences between their signals remains challenging. Existing methods often overlook intra-class similarity and inter-class heterogeneity, particularly within key frequency bands, thus limiting model generalization. To address these limitations, this study proposes a novel contrastive learning framework comprising four core modules: a dual attention mechanism, a dual Mamba module, a classifier, and an advanced contrastive representation learning strategy. Two independent self-attention branches extract diverse and complementary representations, facilitating a degree of feature decoupling through contrastive learning. Each Mamba module applies a contrastive loss to model class-specific similarities and heterogeneities within the learned features. The classifier performs coal-gangue identification by integrating and analyzing the fused feature representations. Ablation and comparative experiments on the simulation dataset validate the method’s effectiveness, achieving 95.03% accuracy, 94.95% F1-score, 94.78% precision, 95.16% recall, and Kappa and MCC scores of 89.89% and 89.94%, respectively, outperforming all baselines. This work provides a robust and generalizable solution for coal-gangue recognition, contributing to the development of reliable, efficient, and autonomous mining systems.