Accurate seismic source location is crucial for assessing rock stability and mitigating dynamic hazards in underground mining. However, high-accuracy source location remains challenging in environments with sparse geophone networks, which are typical in mining operations. The framework strategically integrates three core components in a ‘region-first, precision-later’ workflow: (1) deep learning based on multi-scale attention mechanisms for coarse region location, (2) the regulated signature distance (RSD) method for physics-guided regional refinement, and (3) the sparrow search algorithm (SSA) for metaheuristic-optimized velocity model inversion and grid search. A boundary discrete network is utilized to optimize the initial velocity model structure. The method also integrates a double-difference location technique to reduce velocity model errors. Through comparative and ablation experiments involving both simulations and field tests in coal mines, the method demonstrates superior performance, achieving significantly lower epicenter and source errors compared to traditional approaches. The proposed method distinguishes itself by offering a comprehensive, integrated solution that addresses data scarcity, model inaccuracy, and computational efficiency simultaneously in sparse network scenarios, a challenge not fully tackled by existing hybrid approaches.
Liu et al. (Tue,) studied this question.