Abstract
Time-depth conversion is a crucial step in 3D seismic interpretation of coalfields. Fast and accurate time-depth conversion is essential for ensuring safe and efficient coal production. However, conventional methods often struggle to balance accuracy with convenience, which makes it difficult to achieve good application results in the coalfield. To address this problem, we proposed a new coal seam time-depth conversion method based on machine learning and seismic velocity inversion. Firstly, a high-precision time-domain layer of the coal seam floor was obtained. Subsequently, the average velocity of the coal seam floor was calculated from boreholes. Following this, post-stack seismic inversion was performed to obtain velocity volumes, and the velocity volumes were subjected to median filtering. Next, machine learning models were trained using the average velocity of the coal seam floor extracted from inverted velocity volume, the average velocity of the coal seam floor calculated and interpolated by control boreholes, and two-way travel times (TWTs) of the coal seam floor as inputs, with actual coal seam floor elevations as the outputs. Finally, different machine learning methods and conventional methods were compared and analyzed for time-depth conversion in coalfield. The results indicate that the Bayesian-SVR model achieved the highest accuracy in time-depth conversion, with a maximum absolute error of only 1.11 meters and a mean absolute error of 0.53 meters at verification boreholes. In summary, this study introduces a machine learning-based coal seam time-depth conversion method that does not require complex velocity models, enhancing efficiency while maintaining high accuracy, which holds significant importance for advancing intelligent coal mining and achieving transparent working faces.