Instruments and Systems: Monitoring, Control, and Diagnostics Annotation << Back Creation of a Synthetic Video Sequence for Training Anomaly Detection Systems at Automated Control System Facilities E.V. BORODOV, A.G. GORYUNOV A method for creating a physically sound synthetic video sequence for training and validating hybrid anomaly detection systems in images is proposed and implemented. The approach is aimed at application in the conditions of energy and nuclear power facilities, where the lack of labeled data associated with the development of defects over time is a key limitation in the implementation of automated quality control systems. The main stages of the method are the creation of a standard, modeling the development of defects, adding realistic industrial noise, stress testing, automatic time marking. To create an image norm, the construction of a median frame based on real images and combined alignment is used. Modeling of the development of defects is carried out through a stochastic process model, a model of heat diffusion, textural changes, as well as imitation of shaking and glare. Realistic noise is based on centered Poisson and correlated Gaussian noise. Stress testing is carried out using directional geometric and hardware augmentations. Augmentations perform the following functions: increasing generalizing ability, suppressing overfitting, formal robustness testing, and preparing for tests with remote computing. It is shown that using all stages of the method makes it possible to increase the F1 metric from 0.81 to 0.9, while reducing the frequency of false positives from 22 % to 9 %. The contribution of stress augmentations, which provide resistance to geometric distortions typical of the operating conditions of industrial equipment, is particularly significant. The proposed strategy is implemented as part of data preparation for a hybrid analysis system combining convolutional neural network (CNN) topological data analysis (TDA), wavelets and Shannon entropy. It can be used as part of remote computing systems on a single-board computer, as well as in digital counterparts of critical installations. The result shows the possibility of creating a high-quality, interpretable and reproducible training dataset suitable for training, validation and stress testing of vision systems. Keywords: synthetic video sequence, anomaly detection, robustness, Gaussian noise, Poisson noise. DOI: 10.25791/pribor.4.2026.1662 Pp. 01-10.