Stable neural network control of structurally flexible space manipulators, a novel approach featuring fast training and efficient memory

grantor: University of TorontoAn artificial neural network (ANN) control method is developed for the precision control of elastic-joint robots. Both a novel training method and novel memory structure are developed. The training ensures stabilization by using the Lyapunov-stable technique of backstepping in an on-line, direct neuroadaptive scheme. A learning term is used to modify the on-line weight updates in order to improve performance through learning of the inverse dynamics. The integration of on-line stabilization and learning allows fast learning of the dynamics while maintaining computational simplicity and system stability. Albus's Cerebellar Model Arithmetic Computer (CMAC) memory algorithm is modified to work for elastic systems by utilizing radial basis functions (RBFs) to deal with the elasticity. The resulting hybrid network is referred to as CMAC-RBF Associative Memory (CRAM). Many of the properties of the CMAC for rigid robot control are kept by using CRAM for elastic-joint robots, including computational efficiency and fast convergence. Simulation results are provided to demonstrate the effectiveness of the technique. Two techniques are developed to increase the possibility of real-time implementation. The first is a robust modification. It uses a robust control term to compensate for any shortcomings of the neural network modeling, with increasing torque chatter as the price. Also, a modification to the learning term can be implemented with the robust control that results in better convergence. Simulation results demonstrate the improvement in performance. The second technique is applying the neural network control and weight updates in discrete time instead of continuous time. It is proved that stability can still be achieved. It is shown with simulation results that a larger neural network can be run at a slower (discrete) control rate and performance is improved. The result is a neural control method that is more practical than previous methods.Ph.D