Information-Mathematical Modelling of Machine-Learning-Based Control in Automated Electromechanical Systems

Abstract

In the paper, the quality of adaptive control in a laboratory electromechanical installation operating under variable load and temperature conditions is analysed through a head-to-head comparison of four approaches: a proportional – integral – derivative controller, a fuzzy-logic controller, gradient boosting on decision trees and a long- and short-term memory recurrent neural network. From the information-mathematical point of view, the control loop is formalised as a discrete-time state-space model with a performance functional that jointly penalises tracking error and control effort, and as an information-processing pipeline that combines a digital twin, a Supervisory Control and Data Acquisition (SCADA) system and embedded machine-learning-based controllers. Investigated is the construction of a representative data set of vibration, temperature and pressure with one-second sampling and its integration into an industrial digital-twin environment that reproduces load and thermal stress scenarios. Identified is the impact of stress profiles on mean absolute error, root mean square error, stabilisation time and relative reduction of specific energy consumption as key indicators of control quality and operating costs. Studied is the agreement between the digital twin and the physical installation in terms of correlation, relative error and computation cycle time sufficient for real-time control in industrial conditions. Determined is the statistical significance of differences among all four control strategies using Fisher’s variance test and Student’s t-test for means under multiple disturbance scenarios. Established is that the recurrent architecture provides the most favourable balance of accuracy, transient response and energy saving under non-stationary conditions of modern automated processes. Additionally, an algorithmic scheme is proposed for online training of the long- and short-term memory network and for its deployment on edge devices using high-level programming tools (Python, TensorFlow, Node-RED), which links the mathematical model to a practical information technology solution for industrial control. Formulated is a practical guideline linking improvements in prediction accuracy to reductions in specific energy consumption, which supports management decisions on the selection of intelligent controllers in the context of digital transformation of industrial enterprises.