Abstract:

To improve convergence speed and prediction accuracy of traditional parameter inversion methods in surface deformation monitoring of mining areas, a deep learning framework with physical mechanism constraints was constructed to achieve accurate estimation of surface movement parameters of the working face’s strike and dip in mining areas. Methodologically, based on the theoretical basis of the subsidence prediction model using the probability integration method, the Trend-Net network for the inversion of the 4-parameter strike and the Tendency-Net network for the inversion of the 6-parameter dip of the working face were respectively constructed. The loss function was constructed based on the root mean square error between the predicted value and the measured value of the surface movement, and gradient optimization was carried out to dynamically correct the predicted parameters of subsidence. The experimental results show that compared with the least square method, the particle swarm optimization algorithm, and the Bayesian algorithm, the number of convergence iterations in the inversion of the strike parameter is significantly reduced, and the root mean square error in the inversion of the dip parameter is reduced. This method combines the nonlinear fitting ability of the deep learning network with the physical constraints of the probabilistic integration method. It not only ensures the theoretical rationality of the inversion process but also enhances the global nature of parameter optimization.

Key words: probability integration method, subsidence prediction, deep neural network, parameter inversion, working face of mining area