综采智能化是煤矿智能化发展的关键，对于提升煤矿安全生产水平、实现煤矿少人化及无人化具有重要意义。在液压支架的支撑掩护下，综采工作面采煤机割煤和刮板输送机输煤，二者协同运行形成工作面煤流。由于综采工作面地质条件复杂多变，以及采运作业协同性不足，易造成煤块大小和煤流体积的波动，导致煤流状态异常。因此，对综采工作面煤流异常状态进行智能识别与处理，是保障煤流顺畅、提高综采效率和实现综采智能化的关键。综采工作面常见的煤流异常状态包括：异常煤块和煤流体积异常，目前对上述两种异常状态缺乏有效的智能识别与处理方法。为此，本文依托国家重点研发计划项目“煤矿智能开采安全技术与装备研发”子课题“大采高工作面智能开采安全技术集成与示范”（课题编号：2017YFC0804310），以综采工作面煤流异常状态智能识别与处理方法为研究主线，主要研究综采工作面异常煤块的在线识别和煤流体积的在线测量，以及这两类异常的智能化处理方法，为综采工作面的智能化生产奠定基础。具体内容如下：

（1）提出综采工作面煤流异常状态智能识别与处理系统架构。分析综采工作面采运协同过程和煤流系统及其特征，研究综采工作面煤流异常状态的形成原因，讨论煤流异常状态对输煤顺畅性的影响。从智能感知、智能决策、控制执行这三个关键环节出发，提出一种面向综采工作面的煤流异常状态智能识别与处理系统架构，设计煤流异常状态智能识别模型和处理方案，具体包括异常煤块智能识别、煤流体积异常智能识别、异常煤块智能处理和煤流体积异常智能处理，确保综采工作面煤流畅通。

（2）在综采工作面异常煤块智能识别方法研究方面，提出一种高效的异常煤块识别算法（YOLO-Ghost Representation, YOLO-GR）。YOLO-GR算法在YOLO模型中引入轻量化卷积模块和Vision Transformer模块构建异常煤块特征提取网络。为改善模型在高噪声条件下的特征提取能力，YOLO-GR算法基于参数量解耦的思想，融合RepConv模块和无参注意力机制模块提高模型特征融合效率。依据不同综采工作面地质条件、采高和采运设备的主要参数，制定拟破碎煤块判定标准。在异常煤块数据集上进行对比和消融实验，验证YOLO-GR异常煤块智能识别算法的有效性。

（3）针对综采工作面煤流体积异常识别方法研究，搭建一种基于双目散斑结构光的煤流体积在线测量系统；采用基于霍夫变换直线检测算法的煤流图像分割方法，对煤流图像上刮板输送机中部槽承载的动态煤流区域进行分割；进而采用SGM算法对动态煤流区域散斑图像进行立体匹配，以获得煤流轮廓的点云数据；提出平面密度点云简化算法（Planar Density Simplification, PDS），对煤流点云数据进行简化，通过微元法计算得到煤流体积；提出煤流体积异常判定方法，对煤流体积异常进行判别。通过实验室和煤矿综采工作面现场对煤流体积测量与异常状态识别方法分别进行验证。

（4）针对综采工作面异常煤块智能处理方法的研究，重点解决破碎机器人在线自适应行为决策和异常煤块智能在线处理两个问题。在近端策略优化算法的基础上，引入大核金字塔特征提取机制（Large Kernel pyramid, LK）和双门控注意力机制（Double Gated attention, DG），提出一种具有在线自适应能力的深度强化学习破碎机器人决策控制算法（Large Kernel pyramid Double Gated attention, LKDG）；根据实验环境，给出异常煤块破碎判定依据；在构建的综采虚拟环境中，对LKDG破碎机器人控制算法进行训练；通过仿真实验，验证LKDG异常煤块处理控制方法的有效性。

（5）针对煤流体积异常处理呈现出的高度复杂、动态不确定特征，构建一种基于时空注意力双向感知的异常煤流处理基决策器（Transformer Current attention–BGRU, TC-BGRU），通过与GRU、BGRU、Transformer三种基决策器进行对比，验证TC-BGRU基决策器优劣性；通过对TC-BGRU基决策器进行分层集成，建立煤流体积异常处理动态选择性集成决策模型（Dynamic Integration-TC-BGRU, DI-TC-BGRU），进而提出基于DI-TC-BGRU模型的综采工作面煤流体积异常处理方法；在神东矿区某煤矿52604综采工作面采集生产数据，对DI-TC-BGRU煤流体积异常处理方法的有效性进行验证。

（6）综采工作面煤流异常状态识别与处理方法验证。以煤矿综采实验室三机设备为基础平台，研制综采煤流异常煤块破碎机器人，设计搭建综采煤流异常煤块识别与破碎实验平台，对异常煤块智能识别与处理方法进行实验验证；在神东矿区某煤矿52604工作面，搭建综采煤流体积异常识别与处理实验平台，对综采工作面煤流体积异常智能识别与处理方法进行实验验证。

本文提出的综采工作面煤流异常状态智能识别与处理方法，为综采工作面煤流异常状态的识别与处理提供了新方法，为实现综采工作面智能化奠定了基础。

The intelligentization of fully mechanized mining is the key to  the development of coal mine intelligence, which is of great significance for enhancing the level of coal mine safety production and realizing less manned and unmanned coal mine. The fully mechanized coal mining working face is supported and protected by the hydraulic support. The shearer cuts the coal and the scraper conveyor transports the coal. The two work together to form the coal flow in the working face. Due to the complex and changeable geological conditions of fully mechanized mining faces and insufficient coordination of mining operations, it is easy to cause fluctuations in the size of coal blocks and coal flow volume, leading to abnormal coal flow conditions. Therefore, intelligent recognition and processing of abnormal coal flow conditions in fully mechanized mining faces is the core to ensuring smooth coal transportation and safe production in fully mechanized mining faces. It is also the key to improving the efficiency of fully mechanized mining and realizing the intelligentization of fully mechanized mining. Common abnormal coal flow conditions in fully mechanized mining faces include abnormal coal blocks and volume. Currently, there is a lack of effective intelligent recognition and processing methods for the above two abnormal states. To this end, this article relies on the sub-project of the National Key R&D Plan Project “Intelligent Mining Safety Technology and Equipment Research and Development for Coal Mines” “Intelligent Mining Safety Technology Integration Demonstration for Large Mining Height Working Faces” (Project Number: 2017YFC0804310), based on the coal flow of fully mechanized mining faces. Abnormal state recognition and intelligent processing methods are the main line of research, mainly studying the online recognition of abnormal coal blocks and online measurement of coal flow volume in fully mechanized mining faces, as well as the intelligent processing methods of abnormal coal blocks and coal flow volume abnormalities, which provide the basis for fully mechanized mining work. Laying the foundation for comprehensive intelligent production. The details are as follows:

(1) This dissertation proposes an intelligent recognition and processing system architecture for coal flow abnormality in a fully mechanized coal mining face. This paper analyzes the coordinated process of mining and transportation on the comprehensive mining face, as well as the coal flow system and its characteristics, studies the causes of abnormal coal flow states on the comprehensive mining face, and discusses the impact of abnormal coal flow states on the smoothness of coal transportation. Starting from the three critical links of intelligent perception, intelligent decision-making, and control execution, this dissertation proposes an intelligent recognition and processing system architecture for abnormal coal flow conditions oriented to fully mechanized mining faces and designs an intelligent recognition model and processing plan for abnormal coal flow conditions, including Intelligent recognition of abnormal coal blocks, intelligent recognition of abnormal coal flow volume, intelligent processing of abnormal coal blocks and intelligent processing of abnormal coal flow volume ensure smooth coal flow in fully mechanized mining faces.

(2) In terms of intelligent recognition of abnormal coal blocks in fully mechanized mining faces, an efficient abnormal coal block recognition algorithm (YOLO-Ghost Representation, YOLO-GR) is proposed. The YOLO-GR algorithm introduces a lightweight convolution module and a Vision Transformer module into the YOLO model to build an abnormal coal feature extraction network. To improve the feature extraction capability of the model under high noise conditions, the YOLO-GR algorithm is based on the idea of parameter decoupling and integrates the RepConv module and the parameter-free attention mechanism SimAM to improve the efficiency of model feature utilization. Based on the geological conditions of different fully mechanized mining faces, mining heights and main parameters of mining and transportation equipment, the criteria for determining coal blocks to be broken are formulated. Comparison and ablation experiments were conducted on the abnormal coal block data set to verify the effectiveness of the YOLO-GR abnormal coal block intelligent recognition algorithm.

(3) To study the abnormal recognition method of coal flow volume in fully mechanized mining working face, this dissertation builds an online coal flow volume measurement system in fully mechanized mining working face based on binocular speckle structured light; this dissertation adopts the coal flow image based on Hough transform linear detection algorithm The segmentation method is to segment the dynamic coal flow area carried by the middle trough of the scraper conveyor on the coal flow picture; then use the SGM algorithm to perform three-dimensional matching on the speckle image of the dynamic coal flow area to obtain the point cloud data of the coal flow outline; proposed The Planar Density Simplification (Planar Density Simplification, PDS) algorithm simplifies the coal flow point cloud data and calculates the coal flow volume through the micro-element method. A coal flow volume abnormality determination method is proposed to recognize the coal flow volume abnormality. The coal flow volume measurement and abnormal state recognition methods were verified in the laboratory and on-site at the fully mechanized coal mine working face.

(4) To realize the intelligent processing of abnormal coal blocks in fully mechanized mining faces, this paper focuses on the online adaptive behavior decision-making of the crushing robot and the intelligent online processing of abnormal coal blocks. Based on the proximal strategy optimization algorithm, the Large Kernel pyramid feature extraction mechanism (Large Kernel pyramid, LK) and the Double Gated attention mechanism (Double Gated attention, DG) are introduced to propose a deep reinforcement with online adaptive capabilities, this dissertation presents the decision-making control algorithm of the crushing robot (Large Kernel pyramid Double Gated attention, LKDG). This dissertation proposes to provide criteria for determining abnormal coal block crushing based on the experimental environment. In the constructed virtual environment of fully mechanized mining, the LKDG crushing robot control algorithm is trained. Finally, Verify the effectiveness of the LKDG abnormal coal block processing control method through simulation experiments.

(5) Considering the highly complex and dynamic uncertainty characteristics of coal flow volume abnormality processing, a base decision maker based on spatiotemporal attention bidirectional perception (Transformer Current attention-BGRU, TC-BGRU) was constructed. The advantages and disadvantages of the TC-BGRU base decision maker were verified by comparing it with three base decision makers: GRU, BGRU, and Transformer. Through hierarchical integration of the TC-BGRU base decision maker, a dynamic selective integrated decision-making model for processing coal flow volume anomalies (Dynamic Integration-TC-BGRU, DI-TC-BGRU) was established. Subsequently, a method for dealing with coal flow volume anomalies in fully mechanized mining faces based on the DI-TC-BGRU model was proposed. The effectiveness of the DI-TC-BGRU coal flow volume abnormality processing method was verified using production data collected from the 52604 fully mechanized mining working face in a coal mine in the Shendong Mining District.

(6) Verification of Recognition and processing methods for coal flow abnormality in fully mechanized mining faces. Based on the three-machine equipment of the coal mine’s fully mechanized mining laboratory, a robot for crushing abnormal coal blocks in fully mechanized coal flow was developed. An experimental platform for recognizing and crushing abnormal coal blocks in fully mechanized coal flow is designed and built to conduct experiments on intelligent recognition and processing methods of abnormal coal blocks. At the 52604 working faces of a coal mine in the Shendong mining area, an experimental platform for recognition and processing of abnormal coal flow volume in fully mechanized mining is built to conduct experimental verification of the intelligent recognition and processing method of abnormal coal flow volume in fully mechanized mining faces.

The intelligent recognition and processing method of abnormal coal flow status in fully mechanized mining faces proposed in this dissertation provides a new intelligent method for the recognition and processing of abnormal coal flow status in fully mechanized mining faces, and lays the foundation for the realization of intelligentization of fully mechanized mining faces.