<p>综采工作面智能化是煤炭安全高效开采的发展方向与必然趋势。在智能综采工作面生产过程中，采煤机、刮板输送机、液压支架的协同配合构成了一个多智能体闭环系统。其中，推溜与移架的协同控制不仅需要实现综采工作面连续推进，而且需要调直综采工作面确保满足&ldquo;三直两平&rdquo;的工艺要求以保证煤炭安全高效开采。而如何获取调直综采工作面的有效依据，成为亟待解决的关键问题。综采工作面空间特征主要指煤壁与顶板的交界线等点线面及液压支架等目标体位置信息，点云具有三维坐标可以更全面以数字化的形式表达综采工作面空间特征，因此能够以点云信息为基础对综采工作面空间特征进行解析。</p> <p>本研究以综采工作面点云为研究对象，重点研究综采工作面点云分割模型、分割综采工作面点云的非结构化及无序性处理方法、综采工作面目标体点云识别模型、识别综采工作面目标体点云的稀疏性及几何特征不明显处理方法，从而为综采工作面的自动调直提供关键依据与前提，具体内容如下：</p> <p>（1）分析综采设备协同作业过程，明确综采工作面调直的关键，深入研究综采工作面物理环境下点云所具有的特点，指出基于点云信息解析综采工作面空间特征所面临的难点，进而给出综采工作面空间特征解析的总体思路，为后续研究奠定基础。</p> <p>（2）构建综采工作面点云分割模型感知综采工作面的点线面信息，是以点云形式实现综采工作面空间特征解析的首要内容。首先提出对海量综采工作面点云的快捷标注方法，并高效训练基于动态图卷积神经网络（DGCNN）的综采工作面点云分割模型；其次分析DGCNN中K近邻图算法感受野、多层感知机抽象上升点云特征维数对综采工作面点云不同分割结果的本质原因；最后开展不同前沿深度神经网络的模型对比实验，结果表明：适当选取感受野及点云升维下的DGCNN表现出更好性能，验证了所建立的综采工作面点云分割模型的合理性。</p> <p>（3）针对分割综采工作面点云中点之间的非结构化、无序性特点，在建立综采工作面点云分割模型的前提下，提出多尺度边卷积与串联池化全面满足上述特点。首先构建多尺度边卷积将综采工作面点云的局部特征拼接起来，解决单尺度边卷积只能获取非结构化点云内中心点附近固定邻域结构的问题；其次将串联池化作为对称函数提取综采工作面点云的全局特征，弥补单独使用最大池化满足点云无序性而造成一定特征损失的缺陷；最后开展综采工作面点云分割对比实验，结果表明：集成多尺度边卷积和串联池化的DGCNN（MS&amp;MA-DGCNN）取得最优效果，验证了所提出方法的正确性。</p> <p>（4）构建综采工作面目标体点云识别模型，为综采工作面三维场景提供一个更易于理解的空间世界，更重要的是，为控制液压支架调直综采工作面提供前提。首先制作包含随机噪声的多样性目标体点云（DTPC）数据集，利用DGCNN学习DTPC的特征从而在综采工作面点云中识别目标体点云；其次在改进DGCNN的优化器之后，明确DGCNN在不同边卷积层下的性能；最后开展不同前沿深度神经网络的模型对比实验，结果表明：最佳边卷积层、自适应矩估计优化器的DGCNN（3EA-DGCNN）获得到更好结果，验证了所建立的综采工作面目标体点云识别模型的合理性。</p> <p>（5）针对识别综采工作面目标体点云的稀疏性、几何特征不够明显问题，在建立综采工作面目标体点云识别模型的前提下，提出多密度与增加法矢量解决上述问题。首先在DTPC数据集的基础上构建多密度多样性目标体点云（MD-DTPC）数据集；其次萃取MD-DTPC数据集上所有点的归一化法矢量信息，以形成多密度与法矢量的多样性目标体点云（MDNV-DTPC）数据集；最后开展综采工作面目标体点云识别对比实验，结果表明：MDNV-DTPC下多种深度神经网络性能均有提升，3EA-DGCNN取得最优效果，验证了所提出方法的正确性。</p> <p>（6）进一步说明本文所提出的综采工作面空间特征解析方法在煤矿工程领域下的有效性。首先通过综采工作面点云采集系统获取榆家梁煤矿43101、43206综采工作面点云；其次，综采工作面点云分割感知点线面信息中以顶板空顶区（顶板空顶区仍属于顶板的一部分，后简称为：顶板）与煤壁交界线为例，综采工作面目标体点云识别中以液压支架为例，设计实验方案与细节；最后，在所采集的综采工作面点云基础上进行实验验证，所建立的模型分割43101、43206综采工作面点云的准确率为88.69%、88.87%。所建立的模型识别43101、43206综采工作面目标体点云的准确率为91.41%、91.32%，结果表明：本研究所提出综采工作面空间特征解析方法合理可行且具有普适性。</p> <p>本文提出的综采工作面空间特征解析方法不仅为调直综采工作面提供了依据，并且对数字化工作面综采装备智能识别进而实现联动控制具有积极意义。</p>

<p>Intellectualization of fully mechanized mining face is the development direction and inevitable trend of safe and efficient coal mining. In the production process of intelligent fully mechanized mining face, the cooperation of shearer, scraper conveyor and hydraulic support constitutes a multi-agent closed-loop system. Among them, the coordinated control of sliding and moving supports not only needs to realize the continuous advancement of the fully mechanized mining face, but also to ensure that the straightness of the fully mechanized mining face meets the technological requirements of &quot;three straight and two flat&quot;. Ensuring the straightness of the fully mechanized mining face is the key of&nbsp;safe and efficient coal mining.&nbsp;How to obtain the effective basis for straightness control of fully mechanized mining face has become a key problem to be solved.&nbsp;The spatial features mainly refer to the information in the three-dimensional physical world, such as point, line, surface, body and position and so on. The point cloud with three-dimensional coordinates can more comprehensively express the spatial features of the fully mechanized mining face in digital form. Therefore, the spatial features of the fully mechanized mining face can be analyzed based on the point cloud information.</p> <p>This research takes the point cloud of fully mechanized mining face as the research object. The point cloud segmentation model of the fully mechanized mining face, the unstructured and disordered processing method of the point cloud segmentation of the fully mechanized mining face, the target point cloud identification model of the fully mechanized mining face, and the sparsity and inconspicuous geometric features processing method of target point cloud identification of the fully mechanized mining face are mainly studied, so as to provide the key basis and premise for the automatic straightening of the fully mechanized mining face. The specific contents are as follows:</p> <p>(1)&nbsp;The research analyzes the cooperative operation process of the fully mechanized mining equipment, and defines the key to detect the straightness of the fully mechanized mining face. In addition, the research deeply studies the characteristics of the point cloud under the physical environment of the fully mechanized mining face, and points out the difficulties faced in analyzing the spatial features of the fully mechanized mining face based on the point cloud information. And then it gives the overall idea of analyzing the spatial features of the fully mechanized mining face. These lays the foundation for the subsequent research.</p> <p>(2)&nbsp;The construction of point cloud segmentation model of fully mechanized mining face to perceive the point, line and plane information of fully mechanized mining face is the primary content to realize the spatial feature analysis of fully mechanized mining face in the form of point cloud. Firstly, a fast annotation method for massive point cloud of fully mechanized mining face is proposed, and the point cloud segmentation model based on dynamic graph convolution neural network (DGCNN) is trained efficiently; Secondly, the study analyzes the essential reason of the different segmentation results of the point cloud in the fully mechanized mining face caused by the receptive field of K nearest neighbor graph algorithm in DGCNN and the feature dimension of the abstract rising point cloud of the multi-layer perceptron; Finally, the model is compared with other different front deep neural networks, and the results show that the DGCNN under the appropriate selection of receptive field and point cloud ascending dimension shows the best effect, which verifies the rationality of the point cloud segmentation model established in the fully mechanized mining face.</p> <p>(3)&nbsp;In view of the unstructured and disordered characteristics between the points in the point cloud of the fully mechanized mining face, on the premise of establishing the point cloud segmentation model of the fully mechanized mining face, multi-scale edge convolution and series pooling are proposed to fully meet the above characteristics. Firstly, multi-scale edge convolution is constructed to splice the local features of the point cloud of the fully mechanized mining face, to solve the defect that single scale edge convolution can only obtain the fixed neighborhood structure near the center point in the unstructured point cloud; Secondly, the series pooling is used as a symmetric function to extract the global features of point clouds in the fully mechanized mining face, which makes up for the disadvantage of using the max-pooling alone to meet the disorder of point clouds&nbsp;that&nbsp;causes&nbsp;certain feature loss; Finally, comparative experiments of point cloud segmentation in fully mechanized mining face are carried out, and the results show that the DGCNN integrating multi-scale edge convolution and series pooling (MS&amp;MA-DGCNN) achieves the optimal performance, which verifies the correctness of the proposed method.</p> <p>(4)&nbsp;Building the target point cloud recognition model of the fully mechanized mining face provides a more understandable space world for the three-dimensional scene of the fully mechanized mining face. Most importantly, it provides a premise for controlling the hydraulic support to straighten the fully mechanized mining face. Firstly, the diversity target point cloud (DTPC) dataset containing random noise is produced, and method of identifying target point cloud in point cloud of fully mechanized mining face after DGCNN training is proposed; Secondly,&nbsp;after improving the optimizer of DGCNN, the effect of DGCNN under different edge convolution layers is clarified; Finally, the model comparison experiments of different front deep neural networks are carried out, and the results show that the DGCNN with the best edge convolution layer and adaptive moment estimation optimizer (3EA-DGCNN) obtained better results, which verifies the rationality of the target point cloud identification model established in the fully mechanized mining face.</p> <p>(5)&nbsp;In order to solve the problems of sparsity and insufficient geometric features in target point cloud identification of fully mechanized coal mining, multi density and increased normal vector are proposed on the premise of establishing target point cloud identification model in fully mechanized mining face. Firstly, a multi density diversity target point cloud (MD-DTPC) dataset is constructed on the basis of DTPC dataset; Secondly, the normalized normal vector information of all points on the MD-DTPC dataset is extracted to form a multi density and normal vector diversity target point cloud (MDNV-DTPC) dataset; Finally, comparative experiments on target point cloud identification in fully mechanized mining face are carried out, and the results show that the performance of various deep neural networks under MDNV-DTPC is improved and 3EA-DGCNN achieves the optimal effect, which verifies the correctness of the proposed method.</p> <p>(6)&nbsp;It further illustrates the effectiveness of the spatial feature analysis method of fully mechanized coal face proposed in this paper in the field of coal mine engineering. Firstly, the point clouds of the 43101 and 43206 fully mechanized mining faces of Yujialiang coal mine are obtained through the point cloud collection system of the fully mechanized mining faces; Secondly, in the point cloud segmentation perception point line plane information of the fully mechanized mining face, intersection line between the empty roof area and the coal wall is taken as an example (roof goaf area is still part of the roof, hereinafter referred to as: roof), and in the identification of target point cloud of fully mechanized mining face, hydraulic support is taken as an example to design the experimental scheme and details; Finally, based on the collected point cloud of the fully mechanized mining face, the experimental verification is carried out, and the accuracy of the model to segment the point clouds of 43101 and 43206 fully mechanized mining faces is 88.69% and 88.87%, and the accuracy of the model to identify the target point clouds of 43101 and 43206 fully mechanized working faces is 91.41% and 91.32%. The results show that the proposed method for analyzing the spatial features of the fully mechanized mining face is reasonable, feasible and universal.</p> <p>The spatial features analysis method of fully mechanized mining face proposed in this paper not only provides a basis for straightening the fully mechanized mining face, but also has a positive significance for intelligent identification of fully mechanized mining equipment in the digital fully mechanized mining face to realize linkage control.</p>