METHODOLOGY FOR IDENTIFYING DYNAMIC LINKS USING A FEEDFORWARD NEURAL NETWORK

Abstract

The article considers the methodology for identifying automation objects using a direct propagation neural network.This methodology differs from classical methods of parametric identification in that the control object is considered as a “black” box. The methodology for using a neural network for identification is based on the experimental method of determining the time-dynamic characteristics of the control object. This method involves supplying test signals, such as a step or rectangular pulse, to the input of the object. Depending on the type of test signal, appropriate methods for processing the output signal of the control object are selected. As a rule, when a step control signal is supplied, the acceleration curve of the object is recorded, and when a rectangular pulse signal is supplied, the response curve is recorded. The response curve is recorded for objects that do not allow step signals to be supplied to the input of the object.The research considered the determination of the dynamic characteristics of the object in relation to its acceleration curve when a step test signal is supplied. It is assumed that at the initial moment, the control system should be at rest. In the next step, a step action is applied to the input of the control object, and data on changes in its input parameter over time are collected. When studying the dynamic characteristics of the control object, the following requirements must be met: – if a stabilization system is being designed, the acceleration curve should be taken on the outskirts of the operating point of the process; – acceleration curves should be taken both for positive and negative jumps of the control signal; – in the presence of a noisy output, it is desirable to take several acceleration curves with their subsequent superposition on each other and obtaining an averaged curve.