Automation and control of technological processes and productions

Cloud-based ecosystem of cognitive automation for integrated management of the CIP processes in brewing

A. S. Maksimov^a, V. S. Artemyev^b, L. S. Mangusheva^b, Zh. V. Meksheneva<sup>c

a</sup> Russian Biotechnological University (ROSBIOTECH), 11 Volokolamskoye shosse, Moscow, 125080, Russia
^b Plekhanov Russian University of Economics, 36 Stremyannyy lane, Moscow, 115054, Russia
^c Synergy University, 80B Leningradskiy prospekt, Moscow, 125315, Russia

Abstract: The paper presents a cloud–edge cognitive architecture for managing brewery CIP processes. The system is based on a ResNet-CNN and Transformer ensemble operating within an active learning loop and integrated with multi-sensor monitoring ATP bioluminescence, IR fluorescence, and biofilm optical density. Edge nodes provide instant anomaly detection and local control, while the cloud level performs predictive optimization and model retraining. Pilot trials demonstrated reductions in reagent consumption by 29%, water usage by 22%, and energy use by 18%, along with a decrease in control latency to 140 ms and an increase in predictive accuracy to R$^2$ = 0.92, accompanied by a 37% reduction in false alarms. The architecture ensures compliance with sanitary standards and enables a proactive paradigm for CIP cycle management.
Aim. The aim of the study is to develop a cloud–edge ecosystem capable of reducing decision latency in CIP processes to less than 150 ms, cutting resource consumption, and enhancing sanitary reliability under the conditions of high variability in brewing recipes and technological parameters.
Methods. The methodological foundation relied on the theory of distributed multi-agent systems and the principles of active learning. The dataset included 48,000 fouling profiles, incorporating ATP bioluminescence, IR fluorescence, and biofilm optical density. At the edge level, signal preprocessing is performed, an autoencoder generates compact embeddings, and a GRU-based classifier detects anomalies with a reaction time of less than 40 ms. At the cloud level, a hybrid ResNet-CNN and Transformer model predicts cleaning depth and optimizes CIP cycle parameters. SHAP values and Grad-CAM are used to ensure interpretability of decisions. System validation is conducted in accordance with ISO and GOST standards on metrology, cybersecurity, and sanitary compliance.
Results. The experiments confirm stable real-time operation of the ecosystem and compliance with regulatory requirements. Average consumption of cleaning agents is reduced by 29%, water usage by 22%, and energy demand by 18%. Control latency decreased to 140 ms, while predictive accuracy reached R$^2$ = 0.92. The system demonstrates a 37% reduction in false alarms and full fault tolerance under partial data loss. Economic analysis shows a 24.7% reduction in operating costs and a payback period of less than eight months.
Conclusions. The developed cloud–edge cognitive architecture enables the transition of CIP processes from static operation to proactive control. The combination of fast edge modules and predictive cloud models ensures both resource efficiency and strict sanitary compliance.

Keywords: cognitive automation, CIP, active learning, OPC UA, reagent optimization, industrial IoT

UDC: 681.3.06

MSC: Primary 68T07; Secondary 68T20; 93C85; 90B50

Received: 18.07.2025
Revised: 15.08.2025
Accepted: 25.09.2025

DOI: 10.35330/1991-6639-2025-27-5-143-158

Bibliographic databases:

•