Aiming at the problem of low prediction accuracy in tectonic coal thickness prediction, a new method of Extreme Learning Machine (ELM) optimized by Particle Swarm Optimization (PSO) algorithm was proposed for predicting tectonic coal thickness. Firstly, Principal Component Analysis (PCA) was used to reduce the dimensionality of 3D seismic attributes, which reduced the dimension of seismic attributes, and eliminated the correlation among variables. Then, a Hybrid Kernel Extreme Learning Machine (HKELM) model with global polynomial kernel function and local Gaussian radial basis kernel function was constructed, and the kernel parameters of HKELM were optimized by using PSO algorithm. Furthermore, in order to solve the problem of easily falling into the local optimum for the PSO algorithm, the idea of simulated annealing, the inertia weight decreasing with the number of iterations, and the mutation operation based on reverse learning were added to the PSO algorithm, which made it easier jump out of local minimum points and get better results. In addition, in order to enhance the generalization ability of model, L2 regularization term was added based on the kernel function, which could effectively avoid the influence of noisy data and abnormal points on the generalization performance of model. Finally, the improved prediction model was applied to 15# coal seam in the central part of Luonan No.2 mining area in Xinjing Mining Area of Yangquan Coal Mine, and the predicted thickness of tectonic coal in the mining area guaranteed high consistency with the actual geological data. The experimental results show that the prediction error of the prediction model of tectonic coal thickness constructed by using the improved PSO algorithm to optimize HKELM is smaller, therefore the proposed method can be extended to the prediction of tectonic coal thickness in the actual mining area.

The L1 norm of gradient is used as the regularization term in the Total Variation (TV) model which can preserve the edges of the image well. However, it has the staircasing effect in the relatively smooth regions. Using the variable exponent function as the regularization term, the modified model can not only preserve the edges of image as well as the TV model but also decrease the staircasing effect obviously. Simultaneously, the L1 norm of 〖WTHX〗u-〖WTHX〗f was regarded as the fidelity term of the model, which can enhance the ability of image denoising.