【作者】 陈三风；

【导师】 梅涛；

【作者基本信息】 中国科学技术大学 ， 检测技术及自动化装置， 2008， 博士

【摘要】 研究空间微重力环境地面模拟系统、验证空间机器人的捕获性能是空间机器人研究计划的一项重要组成部分。本论文介绍了目前各种空间微重力环境模拟系统,分析比较了各种模拟系统的优缺点。为了验证正在研究中的空间机器人的捕获性能,本论文提出了两种模拟纳卫星在空间微重力环境下漂浮运动状态的实验平台方案。论文仅研究垂直地面方向的子模拟系统,并重点研究子模拟系统的智能控制系统模块。本论文研究的三维气动式空间微重力环境地面模拟系统具有很强的非线性。其中,主动式模拟系统主要表现为比例阀口气体流动非线性、气体压缩性而引起的非线性以及系统摩擦力所引起的非线性。半主动式模拟系统也是一具有强非线性的复杂系统,其主要表现为气缸推力的随机波动、电机滚珠丝杠系统参数摄动引起的非线性以及系统摩擦力因素引起的非线性。由于气体具有压缩性,模拟系统的刚度比较差、固有频率比较低,系统的抗干扰能力比较差。另外,由于摩擦力的存在,这导致系统的稳态精度和动态性能变差。同时,系统在运行过程中参数会发生变化,这也将导致系统的性能变差。以上原因给控制系统设计带来很大的挑战。传统的控制算法都是基于精确的数学模型,而本研究系统无法精确建模,因此需要研究新的控制方案。径向基神经网络具有很强的非线性逼近能力、自组织能力、自学习能力和自适应能力,它已广泛应用于非线性系统辨识和控制中。但是,径向基神经网络是一种新兴的神经网络,还存在很多不足,需要不断对其进行研究和完善。本文通过优化径向基神经网络的学习算法来提高网络的学习精度、收敛速度、稳定性和泛化能力,同时结合传统的控制算法构成混合控制器,并应用于空间微重力环境地面模拟系统中。仿真和实验结果表明,本文提出的控制算法是有效的,系统具有良好的动态响应能力、鲁棒性和自适应能力。本论文的主要研究工作如下:1.本论文采用主动方案来模拟纳卫星在空间微重力环境下的漂浮运动状态,并对该系统的力学、运动学、气体的流量特性进行了分析和实验。研究结果表明,系统只能响应0-4Hz的抓取力,当抓取力信号频率高于4Hz,系统的动态性能很差。对此,本论文采用了一种改进方案,即半主动式微重力模拟系统,并对该系统的力学、运动学做了一定分析和研究。与主动式模拟系统相比,半主动式模拟系统的动态性能大大提高,可以响应频率高于4Hz的抓取力。2.本研究系统的压力传感器受到很强的电磁干扰和机械振动干扰,论文采用了相应适当的信号处理方案,并针对传统Butterworth低通滤波器在响应速度、检波精度、稳定性三者之间的矛盾,重新设置Butterworth滤波器的主导极点。实验结果表明,本论文的改进方案可有效平衡传统Butterworth滤波器三大主要性能之间的矛盾,可以提取压力传感器的有效信号。3.针对径向基神经网络隐层函数的参数设置难题,论文提出了一种基于密度法的RPCCL聚类算法,基于密度法的RPCCL聚类算法可自动、准确、快速地提取输入样本的特征点,即确定输入样本的聚类数。本论文还成功将其应用于确定径向基神经网络的隐含层结构中。4.针对强非线性系统,论文提出了一种基于共轭梯度下降法的混合PSO算法。该混合PSO算法具有很强的搜索能力,并且能够快速、准确定位全局最优点。与标准PSO算法相比,该算法的收敛速度和学习精度可大大提高。论文还将PSO混合算法成功用于优化径向基神经网络权值中。5.论文提出了一种用于径向基神经网络的在线序列学习算法,即基于极限学习机的在线序列算法。本文的在线序列学习算法既可以学习一个接一个的训练样本,也可以学习一块接一块的训练样本。与其它在线序列算法相比,本论文的在线序列算法的学习速度和学习精度可大大提高。6.为了提高径向基神经网络的泛化性,本论文提出了一种基于局部泛化误差的主动学习算法。该算法采用主动学习方式来选取网络的下一个训练样本,充分利用了系统的已有知识,可以更有效地选择特征训练样本点,减少了网络的训练样本量,大大提高了网络的泛化性能。7.针对三维微重力环境地面模拟系统是强非线性系统,提出了基于径向基神经网络的控制方案。对于主动式微重力模拟系统,采用径向基神经网络来逼近和补偿系统的不确定信息,将它作为前馈补偿使跟踪误差快速收敛,并采用滑模变结构控制来消除径向基神经网络的逼近误差及系统不定随机干扰的影响。实验研究结果表明,这个控制方案是有效的。系统具有较好的动态响应能力、鲁棒性和自适应能力;对于半主动式微重力模拟系统,利用径向基神经网络的学习能力来逼近本研究系统,根据径向基神经网络的学习信息在线调整PID控制器的参数,构造一个具有自调整能力的、稳定的自适应控制器。仿真和实验研究结果表明,这个控制方案也是有效的。更多还原

【Abstract】 One of the key parts of the space robot research project is to develop microgravity simulation platform on the ground in order to test the performance of space robot. The existing microgravity simulation systems are introduced and compared in this paper. To test the performance of the developing space robot, two experimental platforms are proposed for simulating NANO six-degree of freedom floating satellite. The paper only copes with the subordinate simulation systems in the vertical direction, and it focuses on the intelligent control parts of the subordinate systems.The microgravity simulation systems are with strong nonlinearity. The nonlinearity of active simulation system is induced by friction force, compressible air, and the proportional valve. The semi-active simulation system is also a complicated system with strong nonlinearity. Its nonlinearity is introduced by friction force, parameters varying and the random pushing force of air-cylinder. Being air compressible, the inherent frequency of simulation system is very low, which makes the system’s anti-disturbance performance much worse. The system’s accuracy and dynamic performances become much worse for nonlinear friction force. Also, the parameters of simulation system are variable, which also worsen the control system. It’s a great challenge to design control systems for such complicated systems. The traditional algorithms are based on accurate models, yet it’s very difficult to develop accurate models for the simulation systems.Radial Basic Function Neural Network (RBFNN) is good at self-learning, self-organization and approximating the nonlinear system. It has been widely used in nonlinear system identification and control. Whereas, RBFNN is newly proposed, it should be improved continuously. Some new algorithms for RBFNN are proposed to improve the network’s accuracy, stability, generalization performance and converging speed. Hybrid controllers based on RBFNN and traditional algorithms are proposed for the simulation systems. The simulation and experimental results show that these control algorithms are effective. The control systems are with good dynamic performance, robustness and self-adaptive performance.The paper focuses on the following parts:1. An active pneumatic servo scheme is adopted to simulate a NANO satellite how to float in the space. The mechanics, kinematics and hydromechanics of the active simulation system are analyzed. And some experiments are carried out on the simulation system. The experimental results show that the active system can performs well when the frequency of grasping force is lower than 4Hz. Yet it performs much worse at the frequency of grasping force higer than 4Hz. To meet the requirement of the hi-frequncy grasping force, an improved scheme is adopted in the paper, which is a semi-active microgravity simulation system. The mechanics, kinematics of semi-active system are analyzed. The dynamic performance of the semi-active simulation system is much better than that of active simulation system.2. The tense sensor is suffered strong electromagnetic interference and mechanic vibration interference. Some proper signal processing means are proposed for the signal of tense sensor. To balance the contradiction between the test precision, stability and dynamic response time of traditional butterworth low pass filter, the key peak points of butterworth filter are reset. The experimental results show that the new butterworth filter performs much better than the traditional one. It acquires the useful signal of tense sensor successfully.3. It’s very difficult to determine the parameters of hidden layer of RBFNN. A new clustering algorithm RPCCL based on samples’ density is proposed in the paper, which can determine the clustering number automatically, rapidly and accurately. It’s also successfully applied in determining the parameters of hidden layer of RBFNN.4. A new Hybrid PSO algorithm based on conjugate gradient-descent algorithm is proposed for strong nonlinear system. The proposed hybrid PSO algorithm is good at global searching. It can also determine the best solution accurately and rapidly. Compared with the standared PSO algorithm, the proposed algorithm converges faster and more accurately, and it performs well in optimizing the weights of RBFNN.6. A new online sequential learning algorithm based on extreme learning machine is proposed for RBFNN. The learning mode of the proposed algorithm is very flexible, which can learn data not only one-by-one but also chunk -by-chunk. Compared with other online sequential algorithms, the proposed algorithm can converge faster and more accurately.7. The generalization performance of RBFNN is analyzed in the paper. To improve its generalization performance, a new active learning algorithm based on local generalization error is proposed. The proposed algorithm takes the localized generalization error as criterion for selecting the next sample. It takes full use of the previous knowledge, which makes the generalization performance of RBFNNmuch better.8. Control algorithms based on RBFNN are proposed for the microgravity simulation system. RBFNN is employed to approximate and compensate the uncertainties of active simlation system. It works as a feed-forward compensator to make the tracking error convergence fast. The sliding mode controller is employed to obtain robustness of the system for random disturbance and approximation error of RBFNN. The experimental results show that the control algorithm is effective. It produces good dynamic performance, sound robustness and good self-adaptive capacity. RBFNN is employed to approximate the nonlinear system of semi-active simulation system, and the parameters of PID controller is adjusted online according to the approximating result of RBFNN. The simulation and experimental results also show that the control algorithm is effective.更多还原