当煤矿井下发生灾害后，救援机器人可以替代救援人员在第一时间进入灾区实施搜救工作，迅速探明现场环境、及时发现被困矿工，为快速救援提供科学依据。然而，灾后的煤矿井下环境异常复杂且未知，因此，要求救援机器人应具备环境识别、规划决策、运动控制等能力，以保证其在煤矿井下顺利完成救援任务。本文提出并构建煤矿蛇形探测机器人系统，针对该机器人的位姿控制所涉及到的关键科学问题进行研究，包括机器人的相对定位、运动姿态的控制、自主越障的最优控制、对所处环境的有效辨识和自主避障等内容，并研制出一台用于灾后煤矿井下环境探测的蛇形机器人，旨在提高煤矿蛇形探测机器人智能控制和局部自主能力。 在分析灾后井下现场环境的基础上，针对探测机器人的功能需求，设计并研制一种多关节蛇形机器人。该机器人机械本体采用正交关节连接，移动机构采用自制叶片轮驱动。控制系统由硬件平台和软件系统组成，硬件平台为三层次分布式结构，由环境感知层、规划决策层和动作执行层构成。软件系统为功能模块化体系结构，包括环境探测系统、定位系统、姿态控制系统、越障避障模块、通讯系统和上位机系统六个功能模块。该蛇形机器人可为论文的理论及方法验证奠定基础。 针对在煤矿井下封闭式复杂环境下的蛇形机器人定位问题，提出利用机器人运动轨迹曲率与航向角进行估计的方法代替传统的对巷道内复杂地面参数估计的思路，建立基于转向时蛇形探测机器人的定位模型。在此基础上，提出一种卡尔曼滤波算法和深度学习算法相结合的煤矿蛇形探测机器人航位推算的定位方法，该方法采用卡尔曼滤波算法对机器人航向角信号中的高斯白噪声进行剔除，再针对航向角信号中来自低频段的各种漂移，建立基于LSTM（Long Short Term Memory）深度神经网络的陀螺仪输出值预测模型，对未来一段时间内陀螺仪的输出值进行预测，从而实现航位推算的相对定位。实验表明，该方法可以实现蛇形探测机器人的航位推算定位，机器人位置的最小相对误差为3.299 10-12cm, 航向角最小误差为 2.173 10-5rad。 针对煤矿井下不平整地面环境的机器人姿态控制问题，运用 D-H （Denavit -Hartenberg）分析法对正交连接的蛇形机器人进行运动学机理分析，构建两连杆三关节机构模型，可避免建立蛇形机器人运动学模型的复杂性。进而提出基于简化 Serpenoid 曲线的改进的运动控制函数方法，引入曲率误差，建立蛇形机器人蜿蜒、伸缩、抬头不同运动姿态的数学模型，推导不同运动姿态时蛇形机器人的偏转角、俯仰角和相对转角的控制函数，并将仿真得到最优参数运用到机器人的运动控制中，以实现适应复杂地面的蛇形机器人运动姿态的控制，为机器人实现自主越障和避障功能奠定理论基础。 针对机器人在自主越障过程中对关节俯仰角如何确定的问题，提出改进的粒子群优化权系数的极限学习机（Particle Swarm Optimization Extreme Learning Machine，PSOELM）位姿控制算法，该算法推导出动态的惯性系数，利用改进粒子群算法优化极限学习机的隐层矩阵权系数，欲达到优化隐层矩阵的目的，以克服传统极限学习机由于隐层节点参数随机选择而导致网络性能达不到最优的缺点。仿真实验表明，PSOELM 算法不仅延续了ELM 快速学习的特点，而且在控制精度、快速寻优特性及稳定性方面比 ELM 更好，可以实现对机器人关节俯仰角的最优控制。 针对在未知环境下机器人环境辨识建模问题，提出基于遗传算法的变结构模糊神经网络（Genetic Algorithm variable structure fuzzy neural network，GAVSFNN）多传感器数据融合算法，该算法将神经网络引入到模糊规则的推理中，建立基于概率论的改进模糊控制规则，通过改变与简化神经网络结构来调整模糊规则库，实现模糊隶属函数的自学习和模糊规则的优先提取。采用遗传算法优化变结构模糊神经网络的学习参数，以克服常规BP 算法易陷入局部最优的缺点，实现参数的快速和全局寻优。实验结果表明，相比较VSFNN算法，GAVSFNN可以获得对环境模型更高的辨识精度，平均误差为2.725 10-3。在此基础上，建立基于知识库的机器人自主避障系统软模型，以产生避障行为命令对机器人的偏转角进行控制，实现自主避障功能。 以研制的多关节蛇形机器人为平台，与论文提出的几种算法结合，组成煤矿蛇形探测机器人实验系统，在西安科技大学实训基地模拟煤矿巷道中进行现场实验，实验表明，本文所提出的理论和方法可为煤矿蛇形探测机器人在灾后井下非结构化复杂环境中的位姿控制提供理论支撑，也为蛇形机器人在其它灾害环境的位姿控制研究奠定基础。

ABSTRACT After the coal mine accident, rescue robot can replace rescuers in the first time to enter the disaster scene to carry out search and rescue work, quickly identify the scene environment, timely find trapped miners, provide scientific basis for rapid rescue. However, the mine environment after the disaster is very complex and unknown. Therefore, rescue robot is required to have the ability of environment identification, planning and decision-making, motion control and so on, in order to ensure the successful completion of rescue tasks in the coal mine. The snake detection robot system in the coal mine is proposed and constructed, the key technologies involved in the pose control of the robot are studied, which including the relative localization of the robot, the control of the motion attitude, the optimal control of the autonomous obstacle surmounting, the effective identification of the environment and the autonomous obstacle avoidance, and a snake robot for environmental detection in underground coal mine after disaster is developed, aiming at improving the intelligent control and local autonomy of the coal mine snake detecting robot. Based on the analysis of the underground environment after the disaster, aiming at the functional requirements of the detection robot, a multi-joint snake robot is designed and developed. The mechanical body of the robot is connected by orthogonal joints, and the mobile mechanism is driven by self-made blades wheels. The control system is composed of hardware platform and software system. The hardware platform is a three-layer distributed architecture, which consists of environment perception layer, planning decision layer and action execution layer. The software system is a functional modularization architecture, which consists of six functional modules: environment detection system, localization system, attitude control system, obstacle surmounting and obstacle avoidance module, communication system and host computer system. The snake robot can lay the foundation for the theoretical and methodological verification of the thesis. Aiming at the problem of snake robot localization in closed complex environment of underground coal mine, the method for estimating the curvature and path angle of robot trajectory is proposed, which replaces the traditional idea of estimating complex ground parameters in coal mine tunnel, and a simple localization model of snake detecting robot based on turning is established. On this basis, a method of location method of dead reckoning for coal mine snake detecting robot based on Kalman filter algorithm and Deep Learning Algorithms is put forward, which adopts Kalman filter algorithm to eliminate the white Gaussian noise in the path angle signal of the robot. Aiming at various drifts from low-frequency stage in path angle signal, a prediction model of gyroscope output value based on LSTM (Long Short-Term Memory) deep neural network is established, which can predict the gyroscope output value in the future period of time, and the relative location of dead reckoning can be realized. Experiments show that the method can realize the dead reckoning of the snake detection robot, the minimum relative error of the robot’s location is 3.299 10-12cm and the minimum error of the path angle is 2.173 10-5rad. Aiming at the robot attitude control problem of uneven environmental ground in coal mine, the kinematics mechanism of orthogonal snake robot is analyzed by using D-H (Denavit- Hartenberg) analysis method, and the model of three-joint mechanism with two connecting rods is constructed, which avoids the complexity of establishing the kinematics model of snake robot. And then, an improved motion control function method based on simplified Serpenoid curve is proposed, curvature error is introduced. The mathematical models of snake robot with different motion attitudes of serpentine, concertina and head raising are established, the control functions of the deflection angle, pitch angle and relative rotation angle are derived with different motion attitudes of the snake robot. The optimal parameters obtained by simulations are applied to the robot’s motion control, and the control of snake robot’s motion attitude adapted to complex ground is realized, it lays a theoretical foundation for the realization of the robot’s autonomous surmounting obstacle and obstacle avoidance function. Aiming at the problem of how to determine the joint pitch angles of robot in the process of the robot autonomous surmounting obstacle, a pose control algorithm is proposed based on improved Particle Swarm Optimization weight coefficient of Extreme Learning Machine (PSOELM), the dynamic inertia coefficient is derived. In order to obtain the optimized hidden layer matrix of the Extreme Learning Machine (ELM), improved Particle Swarm Optimization (PSO) is applied to optimize the weight coefficient of hidden layer matrix. PSOELM overcomes the shortcoming that traditional ELM cannot reach the best performance because of the random selection of the parameters of the hidden layer nodes. The simulation and experiment results prove that compared with the ELM algorithm, PSOELM algorithm not only continues the characteristics of fast learning of ELM, but also has better control accuracy, fast optimization and stability than ELM, and optimal control of robot’s joint pitch angles is achieved. Aiming at the problem of robot environment identification and modeling in unknown environment, a multi sensor data fusion algorithm based on Genetic Algorithm optimization of the Variable Structure Fuzzy Neural Network (GAVSFNN) is proposed. The neural network is introduced into the inference of fuzzy rule, and the improved fuzzy control rules based on probability theory is established. The fuzzy rule base is adjusted by simplifying the structure of the neural network, the self-learning of fuzzy membership functions and the preferential extraction of fuzzy rules are realized. The genetic algorithm is used to optimize the learning parameters of the variable structure fuzzy neural network, which overcomes the shortcoming that the conventional BP algorithm is easy to fall into local optimum, and the fast and global optimization of the parameters is realized. The experimental results show that compared with VSFNN algorithm, GAVSFNN can obtain higher identifying accuracy for environmental model. The average error is 2.725 10-3. On this basis, the soft model of robot’s autonomous obstacle avoidance system based on the knowledge base is established, which generates obstacle avoidance behavior commands to control the deflection angles of the robot, so that the robot can safely avoid obstacles and the function of autonomous obstacle avoidance is realized. Taking the developed multi-joint snake robot as a platform and combining with algorithms proposed in this paper, an experimental system of coal mine snake detecting robot is constructed. Field experiments are carried out in the simulated coal mine tunnel of Xi’an University of Science and Technology Practical Training Base. Experiments show that the theories and methods proposed in this thesis can provide theoretical support for the pose control of coal mine snake detecting robot in the underground unstructured complex environment after the disaster, and also lay a foundation for the research of snake robot’s pose control in other disaster environments.