• A 24-parameter microseismic index system was reduced to 15 key indicators via manifold learning and hybrid feature selection, enabling effective classification of failure responses.
• Four typical acoustic response types were identified from radar plot morphologies, each corresponding to distinct local failure mechanisms and stress conditions.
• Four representative rock mass instability models were proposed for typical failure zones, integrating focal mechanism inversion and numerical simulation.
• The study provides theoretical and methodological support for hazard prediction and structural optimization in deep metal mining.