Instruments and Systems: Monitoring, Control, and Diagnostics Annotation << Back Automation of Monitoring and Prevention of Pollutant Emissions at Oil Refineries Through the Development of a Machine Vision System for Detecting Air and Water Pollution K.D. SKOBELEV, V.G. BLAGOVESHCHENSKY This paper presents a multi-level automated pollutant emission monitoring system for oil refi neries, focused on the early detection of the "most dangerous" yellow smoke (a proxy indicator of H2 S/SO2 and sulfur aerosol leaks) and the segmentation of thin oil fi lms on the water surface. A compact multi-purpose neural network architecture is proposed with a common convolutional baseline for smoke detection and pixel-by-pixel segmentation of oil films in water and a lightweight temporal module for spatiotemporal consistency of predictions. To improve portability and robustness, a hybrid dataset is used, including open-source data and synthetically generated scenes from digital simulations, as well as augmentation and domain adaptation methods. The model is optimized for execution on edge devices with GPUs and integrates with production analytics via standardized interfaces for generating alerts in the SCADA/ICS system. Experimental evaluation at real sites showed a mAP for smoke detection of ≈ 0.83 and an average IoU for film segmentation of ≈ 0.82 (F1 ≈ 0.90), confirming the practical applicability of the approach for operational environmental monitoring and process risk mitigation. Keywords: monitoring, machine vision, yellow smoke, neural network model, oil film segmentation, domain adaptation, APCS. DOI: 10.25791/pribor.11.2025.1626 Pp. 01-08.