【作者】 徐建安；

【导师】 张铭钧；

【作者基本信息】 哈尔滨工程大学 ， 机械设计及理论， 2006， 博士

【摘要】 21世纪是人类研究、开发及和平利用海洋的世纪，随着人类对海洋开发利用的不断增加，能够探测水下环境并且自主完成特定作业任务的水下机器人受到国内外研究机构的广泛重视。作为在复杂海洋环境下工作的载体，自主性及安全性是水下机器人的重要特征，智能控制技术是保障其自主性和安全性的重要基础和核心技术。水下机器人智能控制的内涵包括自主规划、运动控制与状态监控。研究水下机器人自主任务规划、智能运动控制、传感器信息融合以及自主监控技术，对于提高水下机器人的智能化水平和加快工程化应用进程具有重要的意义。水下机器人是具有较强非线性的复杂动态系统，广义预测控制对于模型辨识误差、传感器噪声以及被控系统的时滞和阶次不确定表现出良好的鲁棒性，本文进行了基于广义预测控制算法的水下机器人运动控制的研究工作。从基于牛顿．欧拉方程和神经网络建立水下机器人的动力学模型入手，通过预测模型，根据被控系统的历史信息和未来输入预测其未来输出，采用有限时段滚动优化，并根据被控系统实际输出误差，在线调整预测模型和控制器参数，实现预测模型与控制器的在线调整。为了研究水下机器人的运动特性，推导了位姿向量在固定坐标系和运动坐标系之间的转换矩阵，根据“海狸”水下机器人的体系结构和运动特点，对转换矩阵进行了简化。根据流体中刚体的牛顿-欧拉方程，建立了“海狸”水下机器人艏向和纵向的动力学模型，利用最小二乘法对动力学模型的参数进行辨识及偏差估计，对“海狸”水下机器人所配置的推进器进行了动态性能分析。在分析动态递归神经网络用于非线性动态系统辨识的原理及可行性的基础上，推导了改进的Elman网络的动态BP学习算法。以滑动窗口模式，采用截短学习样本的方法，实现改进的Elman网络的在线学习，并进行了存在白噪声和类阶跃信号干扰情况下的非线性动态系统的在线辨识。提出了应用于非线性动态系统神经网络辨识的并行模型和串并模型相融合的改进的系统输出递归方式，既对被辨识系统有一定的滤波能力，又提高了神经网络系统辨识的收敛速度。在广义预测控制算法方面，对参数未知、参数慢时变以及考虑控制量及控制量变化率受限的线性动态系统进行了广义预测控制的计算机控制研究，对“海狸”水下机器人的动力学模型进行了动态性能分析，利用欧拉差分得到其离散差分方程，基于特殊非线性动态系统可以时变参数线性转化的理论，对具有二次阻力项的水下机器人非线性动力学模型存在白噪声的情况下，进行了艏向、纵向自由度速度、位移方式的广义预测控制研究。以改进的Elman网络作为多步预测模型，提出并推导了神经广义预测控制律的灵敏度导数计算公式，在存在白噪声干扰和类阶跃信号干扰情况下，分别利用具有在线学习功能的和不具有在线学习功能的神经广义预测控制算法进行了非线性动态系统的预测控制研究和控制误差分析。提出了将具有在线学习功能的神经广义预测控制算法应用于水下机器人的运动控制，并进行了计算机仿真研究，由于改进了神经网络系统输出的递归方式，基于神经网络的广义预测控制相对于基于CARIMA模型的广义预测控制的鲁棒性要好。具有在线学习功能的神经广义预测控制的计算机仿真结果表明具有在线学习功能的神经广义预测控制算法能够实现非线性动态系统的预测控制，且控制效果优于不具有在线学习功能的神经广义预测控制。更多还原

【Abstract】 The 21^th century is the century that man investigate, develop and utilize peacefully the sea, with the development of developing and utilizing the sea, more and more researchers apply themselves to the development of autonomous underwater vehicle（AUV） which can explore the underwater circumstance and accomplish the special missions. As a vehicle which works in complicated oceanic environments, automation and safety are its main character. And intelligent control is the key technology to keep AUV autonomous and safe. Intelligent control includes autonomous mission planning, motion control and status monitoring. So, research on autonomous mission planning, motion control and status monitoring for AUV has the important meaning to improve AUV’s intelligence and application.AUV is complicated non-linear dynamical plant, Generalized Predictive Control（GPC） can systematically take into account real plants constraints in real-time, and is robust with respect to modeling errors, sensor noise. In this paper, some research works about AUV motion control based on GPC are carried out. The dynamical models of the AUV are build with Newton-Euler equations and neural networks and used as multi-step predictive model. With predictive model, based on the passed plant status and future inputs and reference outputs, the future inputs can be predicted, the optimization is real-time. With the actual output error of the real plant, the predictive model or the controller parameters are adjusted. So the optimization is close-loop.Based on analyzing the mechanical structure of the "Beaver" AUV, to describe the motion of AUV, the world reference frame and the body reference frame are used. The transformation matrix for position and attitude vector between two reference frames are also presented. And the transformation matrix is also simplified to "Beaver" AUV. Based on Newton-Euler equations, the AUV dynamical model for yaw and surge are build, the parameters of the dynamical model are also identified with least square method, and the identification error is considered. The dynamics of the propeller for the "Beaver" AUV is also analyzed. The on-line learning for Dynamical Recurrent Neural Networks（DRNN） is proposed and realized with sliding window mode, the nonlinear dynamical plant with white noise is identified on-line with the DRNN, fusing the parallel model and series-parallel model, the improved recurrent mode is proposed. This can not only improve the convergence of the DRNN on-line learning but also filter out the noise. And the DRNN with on-line learning is applied to the "Beaver" AUV dynamical model identification successfully.To GPC, considering the unknown parameters or slowly changing of predictive model and the constraints to the inputs, the indirect and direct adaptive GPC to linear dynamical plant are programmed. The dynamics of the "Beaver" AUV is analyzed. Because a specific non-linear dynamical equations can be linearzed with on-line changing parameters, with Euler differencing equation, the non-linear speed and position dynamical model for AUV are controlled with GPC.The modified Elman Neural Networks is used as multi-step predictive model, the derivative to reason the Neural Generalized Predictive Control（NGPC） law is analyzed elaborately, the on-line learning and off-line learning NGPC is realized to control the non-linear dynamical plant, the output error is also analyzed, the on-line learning and off-line learning NGPC is applied to the control for "Beaver" AUV yaw and surge speed successfully. Because the improved output recurrent mode, the neural networks based GPC is more robust than CARMA model based GPC. When the controlled dynamical plant is polluted with slow changing noise, the control effect to on-line learning GPC is better than off-line GPC.更多还原