Neural Network Control And Deterministic Learning Of Flexible Joint Manipulator

With the rapid development of modern industry,the application fields of robots are becoming more and more extensive.The robotic manipulators are often used to perform different tasks such as carrying items or tracking reference trajectories in space.Common robotic manipulators contain rigid ones and flexible ones.Compared with the traditional rigid manipulator,the flexible joint manipulator has significant advantages in terms of reaction speed,control accuracy,and load-to-weight ratio.Due to the high nonlinearities and strong coupling characteristic,the controller design of the flexible joint manipulator is a greatly challenging problem.On the one hand,due to the introduction of elastic dynamics at the joints,the order of the system increases exponentially.When the controller design is carried out by the traditional backstepping method,the dimensional explosion maybe caused by repeated derivation.On the other hand,neural networks are often used to approximate the unknown nonlinear dynamics of the system.Then,there is another challenging problem,that is,how can we acquire and store the unknown dynamic knowledge from the stable neural control process of the flexible joint manipulator,and how can we use the stored experience knowledge to achieve low-energy and high-performance control for the same or similar tasks? Therefore,the study of adaptive neural control and learning problems for flexible joint manipulators has the important theoretical significance and the broad practical application prospects.The main works of this paper are summarized as follows:Firstly,for the flexible joint manipulator system with unknown nonlinear dynamics,the RBF neural networks are used to approximate the unknown dynamics of the system.By combining the dynamic surface control method,the adaptive neural controller is developed to solve the dimensional explosion and ensure the convergence of the tracking error to a small neighborhood of zero.By combining the deterministic learning theory with the system decomposition technique,the knowledge of the system dynamics is expressed and stored in the constant neural networks during the steady-state control process.Then,the neural learning controller is proposed for the same or similar control tasks by using the stored knowledge.The proposed learning controller avoids repeated training and improves the control performance.Subsequently,in order to avoid the use of multiple neural networks to identify the unknown dynamics of the system and reduce the computational complexity of the control algorithm,a state transformation is used to convert the flexible joint manipulator model into a canonical form.A high-gain observer is designed to estimate the unmeasurable states of the converted system.Then,an observer-based adaptive neural controller is proposed to realize the tracking control of the system.By employing the deterministic learning theory,the knowledge of the sole system dynamics is expressed and stored in the constant neural network.Then,the neural learning controller is proposed for the same or similar control tasks by using the stored knowledge.The proposed learning controller avoids the complex verification of convergences of neural weights,reduces the calculation burden,and improves the control performance.Finally,the thesis focuses on the tracking control problem for the flexible joint manipulator system with unknown dynamics and unmeasurable states.The neural state observer is designed to estimate the unmeasurable states.Then,the convergence of the observation error is verified by using Lyapunov stability theory.By employing the estimated states,the RBF neural network is used to approximate the unknown dynamics of the system.The adaptive neural controller is presented to realize the tracking control of the system and ensure the convergence of the tracking error to a small neighborhood of zero.This thesis also carries out simulation studies to show the effectiveness of the proposed control schemes.