Longwall top coal caving is a key mining method for thick coal seam extraction, where accurate coal and gangue recognition directly affects coal quality and mining equipment stability. Existing recognition methods rely heavily on human expertise or supervised learning models, where the former is subjective and the latter requires a large amount of labeled data, limiting its scalability in mining environments. This paper introduces semi-supervised learning for coal and gangue acoustic signal recognition to reduce dependency on labeled data while enhancing classification performance. A coal and gangue acoustic signal dataset is constructed, and contrastive learning, consistency regularization, and generative adversarial networks (GANs) are employed for comparative experiments to identify the optimal model architecture. Furthermore, this study optimizes the loss function by comparing softmax cross-entropy, center loss, cosface, and arcface in coal and gangue classification. Experimental results demonstrate that arcface loss effectively enhances class separation and improves classification accuracy. Ultimately, under limited labeled data conditions, semi-supervised learning achieves comparable or even superior performance to supervised learning. The findings validate the feasibility of semi-supervised learning for coal and gangue recognition, providing a novel approach for intelligent mining.