• Proposes an adaptive window fractal analysis method for gas concentration data, improving local feature detection over fixed window methods.
• Integrates box-counting dimension and variation metrics to establish a cross-scale dynamic warning model for gas outbursts.
• Achieves dynamic threshold partitioning using membership functions and the 3r principle, enabling graded classification of the MGD index.
• Validated at Shoushan #1 Coal Mine, achieving 86.9% warning accuracy and demonstrating enhanced fluctuation characteristics.