Synthesis of Artificial Intelligence Based Adaptive Controllers for Dynamic Objects

摘要

The article suggests a neural network-based method for implementing adaptive control in automatic control systems for trajectorial displacements of technological objects. The analysis of the structures of automatic control systems with complementary correction allows to propose two schemes for the inclusion of a neural network controller in the automatic control loop: (1) synthesis of the autonomous additional component by the neural network controller, implementation of the component in the control signal and organization of its channel; (2) software control correction implemented in the control signal generating device. These structures presuppose the online learning of the neural network. As an error signal used to adjust the weighting coefficients of the neural network, we propose to use a difference signal between the output coordinate of the reference model and the control object. In this research the corrections to the control law, synthesized in the system correction loop, are assumed to be unknown, i.e. we con-sider the reference value for the neural network output signal to be absent. The work outlines approaches to choosing the number of layers, neurons in them, activation functions and the delay time of input signals of the network. We also offer a modification of the gradient training method for the neural network. The re-search results of the correction loop are presented on the example of the robotic device actuator.