【作者】 宋轶民；

【导师】 张策； 马文贵；

【作者基本信息】 天津大学 ， 机械学， 1999， 博士

【摘要】 本文首次将机构运动弹性动力学、现代控制论及神经网络理论相结合,以具有压电陶瓷作动器与电阻应变计传感器的平面弹性连杆机构为研究对象,系统地开展了基于神经网络的弹性连杆机构振动主动控制的理论、方法与实验研究。本文的研究成果包括:1.为满足非线性动态系统辨识与控制的需要,本文结合了静态前馈神经网络与动态反馈神经网络的优点,采用具有隐层自反馈的Elman型动态递归神经网络(DRNN)以实现基于神经网络的弹性连杆机构振动主动控制。2.本文对神经网络的学习算法进行了深入研究,提出了一种学习步长自适应调节的改进学习算法,大大提高了动态递归神经网络的学习速度,并对这种学习算法的收敛性给予了理论证明。3.本文采用修剪法确定神经网络的最佳拓扑结构,以学习误差对隐层神经元输出的偏导数作为神经网络的灵敏度,逐步去除网络中的冗余神经元,以提高动态递归神经网络的泛化能力。4.在系统辨识方面,本文以实验输入输出数据作为训练样本,采用复合辨识方法离线设计了动态递归神经网络辨识器,获得了弹性连杆机构的非线性动力学模型。系统辨识的结果表明,神经网络辨识器的精度明显高于传统的运动弹性动力学理论模型。5.在参数估计方面,本文构造了弹性连杆机构动力学系统的线性化模型,根据连续型Hopfield神经网络(HNN)的优化计算原理,对该线性化模型的状态方程进行了参数估计。此外,本文还提出了一种基于输入输出数据的线性系统状态空间表达式的HNN参数估计方法。6.本文对基于神经网络的非线性动态系统自适应控制方法进行了深入研究,利用实验输入输出数据离线设计了动态递归神经网络开环、闭环控制器,并分别采用基于神经网络的直接自校正控制方法与基于神经网络的间接自适应控制方法对弹性连杆机构实施了在线控制。结果表明,机构的动态响应得到了显著改善,弹性构件输出点的应变峰值下降了50%左右,明显优于国外其他研究所取得的实验控制效果(10% ~ 20%)。7.由于神经网络理论尚不完善,神经网络非线性控制系统的稳定性分析一直没有明确的理论指导。本文由Lyapunov稳定性定理出发,对弹性连杆机构振动主动控制系统的线性化模型进行了闭环稳定性分析,最终证更多还原

【Abstract】 On the basis of Kineto-ElastoDynamics (KED), Modern Control Theory and Neural Networks (NN), this dissertation studies, in a systematic way, the theory and the methods of the NN based Active Vibration Control of Flexible Linkage Mechanisms whose flexible links are incorporated with piezoceramic actuators and strain gauge sensors. Some achievements acquired are as the follows:(1) To meet the demands of the identification and control of nonlinear dynamic systems, the Dynamic Recurrent Neural Networks (DRNN) are applied in this study, which integrate the nonlinear mapping abilities of the Feedforward NN with the dynamic evolution capabilities of the Feedback NN.(2) An improved learning algorithm is proposed for DRNN whose search steps can be regulated adaptively. The learning speed of DRNN is increased remarkably and the convergence of the algorithm is certified theoretically.(3) The Pruning Algorithms are made use of in the dissertation to determine the suitable topology of NN. The partial derivatives of the learning error to the outputs of the hidden neurons are defined as the sensitivities of DRNN. Those redundant neurons are discarded progressively and the generalization capability of DRNN is improved finally.(4) With the help of the experimental samples, a DRNN identifier is trained off-line utilizing the compound identification method. The nonlinear dynamic model is achieved for the experimental mechanism. Simulation results indicate that the DRNN identifier is more accurate than the traditional KED theoretical model.(5) With reference to the parameter estimation problem, a linearized model of the experimental mechanism is established, whose state equation is determined by means of the continuous Hopfield Neural Networks (HNN). In addition, a novel HNN method is put forward to deal with the estimation problem of the state-space expressions of general linear dynamic systems.(6) This dissertation makes a detailed study of the NN based adaptive control of nonlinear dynamic systems. Two types of DRNN controllers are designed using the experimental samples, which are named the open-loop controller and the close-loop controller respectively. The experimental mechanism is controlled on-line by means of the NN based Direct Self-Tuning Control strategy and the NN based Indirect Adaptive Control strategy. The dynamic response of the experimental mechanism is improved significantly. The crest strain of the flexible coupler is reduced 50 percent or so.(7) Taking advantage of the Lyapunov Theorems, a simplified method is proposed in this dissertation to testify the local stabilization of the experimental control system, which can be used for reference in the stability analysis of general NN based nonlinear control systems.更多还原