• Ludmila V. Kulagina, Siberian Federal University (Krasnoyarsk, Russia)
• Eduard A. Shefer, Siberian Federal University (Krasnoyarsk, Russia)

Ensuring industrial safety at thermal power facilities requires effective and reliable early detection of potentially fire-hazardous and emergency situations. This task is complicated by specific operating conditions characterized by a wide range of operating modes, high levels of electromagnetic interference and vibration, and often a limited amount of labeled data suitable for training and verifying monitoring and forecasting systems. The goal of this study is to develop and experimentally validate a comprehensive approach to improving the accuracy and reliability of computer vision systems for monitoring thermal power facility safety based on an active learning methodology. The primary methods utilized include convolutional neural network models for object detection based on the YOLOv8 architecture, a heuristic mechanism for selecting informative data based on an uncertainty metric, and tools for explainable artificial intelligence and fault-tolerant decision-making with expert participation. The scientific novelty of the work lies in the adaptation of Active Learning methods to single-pass detectors for industrial safety tasks, as well as in the integration of a cycle of targeted additional labeling of complex examples with mechanisms for interpretation and hybrid verification of results (which allows for minimizing labeling costs (due to a selective request for "complex" examples); prompt adaptation to new types of anomalies (e.g., non-standard leaks, local overheating)); whereas most publications on the safety of thermal power facilities use classical machine learning methods (SVM, random forests) or pre-trained CNNs without adaptation to the specifics of the industry. In the completed work, experimental studies on a specialized dataset of images of thermal power facilities showed that the use of active learning provides an increase in the mAP@0.5:0.95 indicator by 28.32% and an increase in the detection recall by 7.82% compared to standard learning. The obtained results confirm the potential of the proposed approach for practical application in monitoring and early warning systems capable of ensuring the timely detection of deviations from normal operating parameters of equipment and preventing the development of emergency scenarios.

active learning, convolutional neural networks, YOLOv8, industrial safety, thermal power plants

2026-06-05