Research And Application Of Time-Varying And Derivative Activated Recurrent Neural Networks Based On Neural Dynamics

In recent years,neural networks have attracted much attention due to their excellent data parallel processing capabilities,good real-time and robustness and other advantages.Among them,recurrent neural network is a popular research direction,and neural dynamics is a method to study recurrent neural network,which regards recurrent neural network as a nonlinear dynamical system and uses dynamical system method to analyze its performance Metrics such as the stability and convergence of the network.In addition to the characteristics of nonlinearity,dissipation,and large degrees of freedom,the neurodynamic model also has a complex attractor structure,so it has excellent computing power.In addition,recurrent neural networks are often used to solve time-varying problems in real time,such as solving time-varying matrix equations and time-varying matrix inequalities in real time.In some engineering fields such as control technology,these problems often exist,such as optimization problems,robot motion planning problems,redundant manipulator arm obstacle avoidance problems,UAV control optimization problems,etc.However,traditional numerical algorithms are difficult to meet the real-time requirements for solving these complex time-varying problems.Therefore,this paper investigates neural dynamics-based recurrent neural networks and applies them to the solution of the Sylvester equation,time-varying linear equations,and the design of UAV controllers.The main work of this paper includes the following aspects:First,a varying-factor finite-time recurrent neural network solver is proposed based on neural dynamics,and the time-varying Sylvester equation is solved using this solver.The network achieves finite-time convergence without using a specific activation function.In the theoretical analysis,the proof of the finite-time convergence property of the network under several different typical activation functions,and the proof of the robustness of the network under different noise conditions are given.In the simulation experiment,by comparing with the zeroing neural network,it is verified that the varying factor finite time recurrent neural network has better convergence and robustness.Then,a rapid-convergent recurrent neural network based on derivative term activation is proposed.By activating the derivative term,the network greatly improves the convergence speed.In the theoretical analysis,the convergence is proved for a class of time-varying problems.In the simulation verification,the convergence performance of the proposed rapid-convergent recurrent neural network is verified by comparing with the zeroing neural network and the varying parameter convergent differential neural network.Finally,the height and attitude angle controllers of the UAV are designed by using the varying-factor finite-time recurrent neural network,so that the UAV can complete the tracking task of time-varying targets.The design formula and process flow of the controller are introduced in detail,and then the feasibility of the controller design scheme is verified through simulation experiments,and compared with the UAV controller based on the varying parameter convergent differential neural network to verify that the controller has faster speed and higher precision.