煤炭作为我国能源支柱之一，在国家工业生产和经济发展中扮演着不可或缺的角色。在煤炭开采过程中，准确识别煤岩界面对于合理规划开采方案和确保矿井安全具有至关重要的作用。然而传统的地质勘探方法往往需要耗费大量的时间和人力，并且工作面地质条件复杂多变，煤岩界面的形态和特征变化难以准确把握。为实现矿产资源的高效开采和最大化利用，本文采用深度学习技术对煤岩界面进行准确识别，主要研究内容如下：

（1）针对钻探数据稀疏导致煤层模型较为粗糙的问题，提出一种基于最小二乘支持向量回归优化克里金的煤层三维地质建模方法。通过对已知钻探数据点的空间关系和趋势进行分析，利用插值算法推断未知位置的煤层高度，从而将离散的数据点转化为连续的煤层模型。为提高插值结果的准确性和可靠性，利用最小二乘支持向量回归求解最优拟合结果以精确重构变差函数。运用仿真煤层曲面数据和实际钻孔数据进行实验，实验结果表明，改进后的克里金算法插值精度更高，均方误差值为0.0041，为煤岩界面识别提供了更准确的煤层三维地质模型。

（2）针对采煤机在复杂煤层下自动截割精度较低的问题，提出一种基于注意力机制结合多元深度门控循环神经网络的截割轨迹预测算法。利用截割轨迹纵向及横向相邻数据之间的相关性，通过多元门控循环神经网络提取滚筒高度数据序列中的复杂非线性特征预测截割路径。为提高预测效率以满足循环截割的实时性要求，利用因果卷积提前聚焦序列纵向的局部时间特征提高运算速度，同时通过自注意力机制进一步提升预测精度。实验结果表明，平均绝对误差值为43.49 mm，平均绝对百分比误差值为1.89%，均方根误差值为50.05 mm，预测时间仅为0.17 s，改进后的截割轨迹预测算法能够更准确地对采煤机截割轨迹进行实时预测，为工作面煤层模型的修正提供了可靠依据。

本文通过某矿实际钻探数据融合动态采煤机截割示教数据，首先插值建立初始煤层三维地质模型，然后利用时间序列预测技术预测未知煤层形态对模型进行精细化修正，最终实现煤岩界面的高效、精准识别，为煤炭智能化开采提供关键技术支撑。

As one of the energy pillars in China, coal plays an indispensable role in national industrial production and economic development. In the process of coal mining, accurate identification of the coal-rock interface plays a crucial role in rational planning of the mining scheme and ensuring mine safety. However, traditional geological exploration methods often require a lot of time and manpower, and the geological conditions of the working face are complex and variable, making it difficult to accurately grasp the morphology and characteristics of the coal-rock interface. In order to achieve efficient mining and maximize the utilization of mineral resources, this paper adopts deep learning technology to accurately identify the coal-rock interface, and the main research contents are as follows:

(1) Aiming at the problem of sparse drilling data leading to a relatively rough coal seam model, a three-dimensional geological modeling method of coal seam based on least squares support vector regression optimization kriging is proposed. By analyzing the spatial relationships and trends of known drilling data points, the interpolation algorithm is used to infer the height of the coal seam at an unknown location, thus transforming the discrete data points into a continuous coal seam model. To improve the accuracy and reliability of the interpolation results, least squares support vector regression is utilized to solve the optimal fitting results in order to accurately reconstruct the variance function. Experiments are conducted using simulated coal seam surface data and actual borehole data, and the experimental results show that the improved Kriging algorithm has a higher interpolation accuracy with a mean square error value of 0.0041, which provides a more accurate three-dimensional geologic model of the coal seam for coal-rock interface identification.

(2) Aiming at the problem of the low accuracy of automatic cutting of coal mining machine under complex coal seams, a cutting trajectory prediction algorithm based on attention mechanism combined with multivariate deep gated recurrent neural network is proposed. Using the correlation between the longitudinal and lateral neighboring data of the cutting trajectory, the complex nonlinear features in the drum height data sequence are extracted by the multivariate gated recurrent neural network to predict the cutting path. In order to improve the prediction efficiency to meet the real-time requirements of cyclic truncation, the causal convolution is utilized to focus the local time features in the longitudinal direction of the sequence in advance to improve the operation speed, and the prediction accuracy is further improved by the self-attention mechanism. The experimental results show that the average absolute error value is 43.49 mm, the average absolute percentage error value is 1.89%, the root-mean-square error value is 50.05 mm, and the prediction time is only 0.17 s. The improved cutting trajectory prediction algorithm can more accurately predict the cutting trajectory of the coal miner in real time, which provides a reliable basis for the modification of the coal seam model of the working face.

In this paper, the actual drilling data of a mine is fused with the dynamic coal miner cutting demonstration data, firstly interpolated to establish the initial three-dimensional geological model of coal seams, and then the time series prediction technology is used to predict the morphology of the unknown coal seams to make fine corrections to the model, and ultimately to realize the high efficiency and accurate identification of coal-rock interfaces, which provides a key technical support for the intelligent mining of coal.