2026, Vol. 13, No. 3. - go to content...

Permanent address of this page - https://resources.today/en/15inor326.html

Метаданные этой статьи так же доступны на русском языке

DOI: 10.15862/15INOR326 (https://doi.org/10.15862/15INOR326)

Full article in PDF format (file size: 853.3 KB)

For citation:

Azizmohammadi M.-S.M., Goncharov A.V., Shakhovskoy A.V., Rotanov E.G. Improving the efficiency of technological process control based on intelligent processing of production data. Russian Journal of Resources. 2026; 13(3). Available at: https://resources.today/PDF/15INOR326.pdf (in Russian). DOI: 10.15862/15INOR326

Improving the efficiency of technological process control based on intelligent processing of production data

Azizmohammadi Mohammad-Safar Mohsenovich
K.G. Razumovsky Moscow State University of Technologies and Management, Moscow, Russia
Email: safar2002@yandex.ru

Goncharov Andrey Vitalievich
K.G. Razumovsky Moscow State University of Technologies and Management, Moscow, Russia
E-mail: a.goncharov@mgutm.ru

Shakhovskoy Andrey Vladimirovich
K.G. Razumovsky Moscow State University of Technologies and Management, Moscow, Russia
E-mail: a.shakhovskoy@mgutm.ru

Rotanov Evgeny Gennadievich
K.G. Razumovsky Moscow State University of Technologies and Management, Moscow, Russia
E-mail: e.rotanov@mgutm.ru

Abstract. Distributed control systems of continuous production accumulate large volumes of production data, yet product quality indicators are still measured by on-line analysers with a delay, so operating decisions rely on outdated information. The paper proposes and tests an approach in which the result of intelligent processing of production data is related to the efficiency of technological process control, while model accuracy serves as an intermediate characteristic. The empirical base is an open industrial data set of a refinery debutanizer column comprising seven process variables and the butane content of the bottom product. The task is stated as an anticipatory estimate of the analyser reading at a given horizon from the information available at the moment of decision. A naive scheme, a univariate autoregression and four multivariate models, namely ridge regression, projection to latent structures, gradient boosting and a neural network, are compared under a chronological split of observations and expanding-window cross-validation. Process variables are shown to lead the analyser reading, and their inclusion reduces the error of the anticipatory estimate the more, the longer the horizon, cutting it by about half relative to the univariate scheme at medium horizons. The regularised linear model performs no worse than the nonlinear ones, whereas standard outlier filtering worsens the result and is rejected by cross-validation. The main contribution to accuracy comes from the temperature on the control tray, while some variables are informationally redundant. Based on the known result on the lower bound of the variance of a controlled variable, the error of the anticipatory estimate is interpreted as the limit of attainable stabilisation quality. This makes it possible to express the effect through the reduction of the irreducible variability of the indicator, the admissible shift of the operating point towards the specification limit and the completeness of early warning of threshold exceedance. The results can be used to justify advanced process control projects and to select data for soft sensors.

Keywords: technological process control; intelligent data processing; production data; soft sensor; artificial intelligence; machine learning; measurement delay; quality indicator; distillation