Formation of digital counterparts of automatic control objects using neural networks in cold fish drying processes

Abstract

The paper considers the application of neural networks in the construction of a digital twin of a technological process. The purpose of the presented research is to process the data accumulated by the automatic control system of a small-sized drying plant about the technological processes occurring on it during the last time, to train a neural network on their basis and to form a digital (neural network) model of temperature change in a thermal chamber. Elements of machine learning are used involving a multilayer neural network of direct propagation. The method of error back propagation is used in the work where the error value of the output neuron is projected onto all the weights of all the neurons of the network, starting from the output and ending with the weights of the neurons of the input layer. During the training, the network received information about the power of the plant's actuators and the temperature in the thermal chamber changing over time. Upon completion of the training, the state of the neural network was formed, which is a digital model of temperature changes in the thermal chamber of a small-sized drying plant. The model obtained with the help of a neural network (digital twin) has shown a correlation with experimental data with an average absolute percentage error not exceeding 3 %. Thus, the behavior of the neural network model is adequate to the real object. Further research in the field of the formation of a digital twin of a technological object (taking into account additional parameters in the model, formation of a neuroregulator based on the model) is necessary and planned by the authors.