To address the challenges posed by insufficient illumination, severe dust interference, and multi-scale target deformation in underground coal mine environments, this paper proposes a Multi-Scale Self-Adaptive Fusion (MSAF) module based on the YOLOv11 framework. Specifically, MSAF enhances small-object textures in shallow-layer features through a local magnification mechanism, effectively suppresses dust and smoke interference via an adaptive noise suppression branch, and deeply integrates detailed textures and global semantics from multi-resolution features using a dual-attention fusion strategy. Experimental results demonstrate that MSAF achieves precise detection of both small distant targets and large nearby equipment under extreme mining conditions, significantly improving model robustness and detection accuracy with minimal additional computational overhead. Thus, this method provides feasible technical support for enhancing mine safety and intelligent inspection.