Terminal Neural Networks With Its Applications

The dynamic neural networks are applied to solve the time-varying problems for the efficient parallel processing abilities.Terminal neural networks are applied to deal with time-vary problems.The dynamic processing ability is of asymptotical characteristics.The convergent performances are improved by inducing terminal attraction.In this disserta-tion,the terminal neural networks and terminal computing problems are discussed.New solving proposals are advised for convergence of different time-varying problems.Further-more,calculations of the terminal neural network are used to the trajectory planning with redundant manipulators.The main research work of this dissertation is summarized as follows:1.Four different types of terminal neural networks are proposed based on asymptotic neural network and the convergent stability of the neural networks are given.Computational results show that the terminal neural network model has the characteristics of finite time convergence and can obtain high convergence precision.In order to speed up the convergent rate,different excitation functions are designed in models of the terminal neural network to improve the response time of the dynamic system.2.For the problems of time-varying inverse calculations and time-varying Sylvester matrix equations,four different types of the terminal neural network are constructed and simulation results verify the validity of the terminal convergent characteristics.It is shown that the error convergence can be achieved as the scaling parameter increases,while the variables of the neural system undertaken are bound.3.For the problems of time-varying inequality matrix equations and time-varying gen-eralized inverse matrix equations,four different terminal neural networks are constructed.The convergence analysis indicates that the tracking error achieves in finite time,while asymptotic neural network approach to zero in infinite time.Numerical experiment results demonstrate the effectiveness of the modeling and calculating method.4.For time-varying quadratic programming problems,by analyzing the solving pro-cess,the time-varying equations can be obtained.The problem of the equations are solved with different terminal neural networks.Compared with the calculation results obtained by asymptotic neural network,experiment results demonstrate the superiority of the modeling and calculating method.5.To solve the joint-angle drift problems in cyclic motion of redundant robot ma-nipulators,a kind of quadratic optimization models for redundant manipulators' trajectory planning based on terminal optimality criterion is proposed and analyzed.The terminal neural network models with limited value activation functions are applied to redundant manipulators in performing the repeatable motion planning tasks while the initial position deviates from the target position.A type of terminal neural network(TNN)with its accel-erated form(ATNN)is proposed,which is of terminal attractor characteristics and it can be obtained effective solutions for time-varying matrix problems in finite time.The simulation results show that the proposed method can achieve faster convergent speed and exponential convergence precision.6.The repeatable problem of mobile manipulator robotic systems in the presence of a fixed base is addressed.Redundancy resolution,schemes based on terminal neural networks are proposed by con,sidering each joint angle of redundant manipulators deviate from the desired position.It is shown that the redundant manipulators converges to the desired trajectory over a special time interval,whatever initial value manipulators take.Numerical simulation results are shown feasibility of the proposed schemes.