【目的】采空区智能充填是煤炭资源绿色安全智能高效开采的重要方向,其关键在于对井下采空区矸石充填过程进行智能决策与控制。【方法】为实现该目标,以采空区充填后围岩应力及变形作为监测指标,建立了一种采空区智能充填深度神经网络算法,该算法可以通过输入煤层埋深、厚度、工作面长度、直接顶厚度等关键基本参数,进行相应条件下不同充填方案的采场应力及围岩变形计算分析。将FLAC3D模拟400种不同条件下的充填开采结果作为数据集,对建立的智能充填深度神经网络算法进行训练测试,并和其余3种不同算法进行对比分析。【结果】结果表明:建立的智能充填深度神经网络算法总体优于随机森林、决策树和多元线性回归算法,每组数据运算平均速度仅为0.013s;智能充填深度神经网络算法计算的顶板最大变形、工作面煤壁压力峰值、巷道超前支护距离等关键参数误差均值介于2%~8%之间;应用该算法针对现场实际条件进行测试,结果与现场实际结果基本吻合,表明该算法科学可行。【结论】本研究对煤矿绿色智能开采具有重要意义与价值。

【Purposes】 The intelligent filling of goaf is an important direction of green, safe, intelli-gent, and efficient mining of coal resources, and the key lies in intelligent decision-making and control of gangue-filling process in underground goaf.【Methods】 To realize this, the stress and deformation of surrounding rock after filling are taken as monitoring indicators, and a deep neural network algo-rithm for intelligent filling in goaf is established, which can calculate and analyze the stope stress and surrounding rock deformation of different filling schemes under corresponding conditions by entering key basic parameters such as coal seam burial depth, thickness, working face length, and thickness of the direct roof.  By using simulation results of FLAC3D under 400 different conditions as a dataset, the intelligent filling deep neural network algorithm was trained and tested, and compared with other three different algorithms.【Findings】 The results show that: the intelligent filling deep neural network algorithm is generally better than the random forest algorithm, decision tree algorithm, and multiple linear regression algorithm, and the average caculation speed of each group of data is only 0. 013 s; The average error values of key parameters such as the maximum deformation of the roof plate, the peak pressure of the coal wall of working face, and the advanced support distance of roadway calculat-ed by the intelligent filling deep neural network algorithm are between 2%~8%; The algorithm is tested according to the actual conditions of the site, and the results are basically consistent with the ac-tual results on the site, indicating that the algorithm is scientific and feasible.【Conclusions】 This study is of great significance and value to green and intelligent mining.