Synthesis of neural network controllers for objects with non-linearity of the constraint type

Abstract

The paper discusses the use of artificial neural networks of direct propagation to control objects with nonlinearities such as saturation and a rigid mechanical stop. The multilayer structure of a neural network is considered using the ReLU activation function and the Dropout layer, which allows working with the indicated nonlinearities. The possibility of training a neural network to control piecewise linear objects of a general type is demonstrated. Using the method of generalized inverse neurocontrol, a controller is synthesized that provides an acceptable quality of control in the presence of such nonlinearities. A specific example shows the possibility of creating a direct neuroemulator that repeats the dynamics of nonlinearities such as saturation and a rigid mechanical stop. The work of the neuroregulator in the examples is compared with the PID controller configured in Matlab. All theoretical conclusions are confirmed by numerical modeling of real technical systems with neural network control.