On The Model Identification And Control Of Robots Based On Neural Networks

The modeling and control of robots is one of the most important fields in the study of robots. The artificial neural network has attracted great interests for the learning ability, adaptive capability and nonlinear mapping property. The modeling and control of robots is studied by the method of neural networks in this dissertation. The simple and effective methods are explored to solve the problems of identification and control in the nonlinear system. This study has great significance in theory and application. The work includes:Firstly, the topological structure of neural network is studied. A new structure of neural network named State Delay Input Dynamical Recurrent Neural Network (SDIDRNN) is proposed based on dynamical recurrent Elman network. The learning ability of SDIDRNN is improved by inputting the prior input-output knowledge into the neurons in the last hidden layer. The learning algorithm and convergence of SDIDRNN are discussed intensively. The learning superiority of SDIDRNN is illustrated through comparison between several commonly used networks. This lays the foundation for the whole study.Secondly, the control of robots is studied based on SDIDRNN. Two control systems are designed and applied to the trajectory tracking based on the kinematics and dynamics of robots. The numerical simulation shows that SDIDRNN controller plays an effective role in the systems. The limitation of PD controller used only is discussed. The PID controller is combined with neural network to solve the problem of the PID parameters unable to be tuned online. The neural network tunes the parameters online to make PID controller more adaptive in the system. This method is applied to a dual-arm system manipulating a rigid object coordinately. The joint controllers for the leader and the follower of the coordinated system are designed. Thirdly, a novel method is proposed to identify the forward and inversekinematics of redundant robots. This new decoupling identification scheme is designed to improve the computational speed for kinematic model identification of a spatial 7R redundant robot. The learning ability is improved greatly by adopting variable weights and constant weights in the scheme. Fourthly, the input-output models of robots are approximated by neural networks in the experiment. The input and output knowledge of neural network is provided by the signals of joints returned by PowerCubeTM modular robot and the trajectory of the end effector measured by 3D motion measurement system OPTOTRAK 3020. The input-output models of three types of robot systems are approximated. The factors that effect the experimental results are analyzed. At the same time, the learning superiority of SDIDRNN is approved by comparing it with other commonly used networks.Finally, the neural network models for robots are tested through experiments. SDIDRNN is used to identify the input-output model of the robot and a neural network model representing the real input-output mapping is formulated after training. This model is proved to be valid in experiment by inputting other different samples. The output knowledge of the model is input to PowerCubeTM modular robot. The trajectory of the end effector is close to the desired trajectory. Elman network is also used in the experiments. The experimental results indicate that SDIDRNN is superior to Elman in learning ability. In addition, because the trajectory error of the end effector is increasing when the joints are motivated by untuned joint signals, a tuning method of joints is proposed to reduce the trajectory error. The effectiveness of this method is proved in the experiment.