To address the challenges of strong background noise, scarcity of labeled samples, and insufficient interpretability of multimodal fusion in rolling bearing fault diagnosis under complex underground coal mine conditions, a two-stage diagnostic framework combining Hierarchical Multimodal Contrastive Learning and the Shapley method (HMCL-Shapley) is proposed. In the first stage, a ternary complementary view comprising time-domain, frequency-domain, and gray recurrence plots is constructed to extract latent features from diverse representation spaces. A curriculum learning mechanism is introduced to progressively enhance the anti-noise feature learning capabilities of the model. On this basis, a Hierarchical Collaborative Momentum Contrastive (Co-MoCo) method is designed. Through self-supervised pre-training, both unimodal discriminative features and cross-modal semantic correlations are jointly learned, enabling the model to acquire stable fusion representations and mitigate overfitting risks during supervised fine-tuning with limited labels. In the second stage, the pre-trained model is fine-tuned using a small number of labeled samples. A Shapley value-based fusion decision strategy is integrated to quantify the contributions of individual and fused modalities, whereby decision weights are assigned to improve the reliability and interpretability of classification results. Experimental results on the Paderborn University (PU) dataset demonstrate that the proposed method achieves a diagnostic accuracy of 84.60% in a 1-shot (extremely few labels) setting, representing a 6.59% improvement over baseline methods. Robust identification performance is maintained under low signal-to-noise ratio conditions, validating the efficacy and suitability of the framework in scenarios characterized by label scarcity and high noise.