Sparse crowd sensing method for methane concentration field in coal mine fully mechanized mining face

In view of the lack of methane monitoring points in coal mine fully mechanized mining face, it is difficult to achieve for high coverage methane concentration monitoring in the mining face area. We propose a methane concentration field sensing method in the mining face based on sparse crowd sensing. Through the monitoring data of methane concentration in some sub-cells, the correlation of sub-cells is explored, and the methane concentration field of fully mechanized mining face is inferred. Firstly, based on the sub-cell division of fully mechanized mining face, we propose a deep reinforcement learning based on distributed weighted self-attention mechanism (DSA-DQN) methane concentration sensing cell optimization selection algorithm. The cell optimization selection algorithm uses the distributed weighted self-attention mechanism (DSA) to capture the key information of the mining face environment, to assist the optimization and decision-making of the cell optimization selection algorithm based on DQN. Secondly, in the aspect of methane concentration inference in the unsensing area, a Kriging interpolation based on particle swarm optimization support vector regression (PS-Kriging) methane concentration inference method is proposed. The support vector regression is combined with the Kriging interpolation method to solve the problem that the traditional method fall into the local optimal solution in the interpolation process. Particle swarm optimization is used to improve the kernel parameters and penalty parameters of the support vector regression model, which improves the inference accuracy of the model. Finally, we choosed a fully mechanized mining face in a Shanxi province mine, and made experimental verification through the combination of on-site measurement and simulation experiments. The experimental results show that the sparse crowd sensing method can infer the methane concentration in all cells of the mining face by selecting some key sub cells, with an mean absolute error of 0.07%.