Designing A Robotic Manipulator's Controller With Recurrent High Order Neural Networks

It is widely accepted that neural network provides an effective alternative in system identification and controller design. In this thesis, a dynamic recurrent high-order neural network is considered which features a straightforward architecture and small neurons. This network is designed and implemented to tuning and tracking a 2-DOF robot manipulator.In the opening part, the basic background information is presented first, such as some historical notes of neural network, the necessity of using dynamic recurrent neural networks and some primitives of robot manipulators.Then neural networks are shown the potential of approximate ability, robustness in functioning and adaptability in dealing multi-variable systems. Till today, neural networks have grown to be an essential tool in system identification and adaptive control. After the architecture of the neural network is chosen, the procedure of identification can be viewed as optimizing some criterion function by adjusting the values of those weights according to the pre-observed data of the plant to be identified. Static neural networks have some shortcomings in model identification.One dynamic recurrent high-order neural network's architecture is applied. This network can be used to obtain plants'dynamic features with a high bandwidth. Here the plant is a robot manipulator which features high-coupling and highly nonlinearity. Based on Lyapunov stability theory, on-line learning rule is derived. This rule's global stability is guaranteed and its global approximating ability is prominent. Thus it solved the difficulty in designing controllers with the presence of un-modeled dynamics and external disturbances.The ultimate goal of identification is to control the plant. After successful identification, a controller based on this dynamic recurrent high-order neural network is formulated further. Thus this network's role is two-fold: an on-line identification component and controller constructor. Simulation shows that this network architecture and the weight adjusting algorithm formulated here did a nice job in tuning and tracking of 2-DOF robot manipulator.