Abstract: The complex environment of mining working faces—including dust, high humidity, and smoke—causes severe feature degradation in monitoring images under varying fog concentrations. Moreover, existing dehazing models trained mainly on synthetic data exhibit domain gaps with real mining fog, limiting intelligent monitoring effectiveness and posing safety risks. This study proposes a dehazing method for working face images based on fog grading and domain differences. First, fog evaluation metrics guide image grading, enabling adaptive network selection for light and dense fog scenarios. Second, a contrastive learning strategy refines negative samples based on fog concentration, improving feature discrimination and cross-domain generalization. Finally, an unsupervised fine-tuning strategy with cyclic consistency mitigates domain bias between synthetic and real fog images without requiring annotations. Experiments show that the proposed method outperforms existing approaches on both synthetic and real datasets, supporting safe and intelligent monitoring in coal mines.