Predictive modeling in computer vision: hybrid deep learning architectures for multidimensional time series analysis

The relevance is determined by the key problem of modern intelligent systems — the limited effectiveness of existing computer vision models for predicting dynamic processes in changing conditions. The need to develop high-tech technologies capable of analyzing multidimensional time series and ensuring high reliability of forecasts is a fundamental challenge in the field of artificial intelligence. The aim of the research is to develop methodological foundations for creating predictive computer vision models based on hybrid deep learning architectures for time series analysis of multimodal data. Methods. Hybrid deep learning architectures, including combinations of convolutional (CNN) and recurrent neural networks (LSTM), as well as ConvLSTM models, have been investigated. A mathematical apparatus has been developed to formalize learning processes, including specialized loss functions and optimization algorithms adapted for forecasting tasks based on time series of visual data. Results. It has been experimentally shown that the CNN-LSTM hybrid model predicts the dynamics of the development of objects 8 time steps ahead with an average absolute error of 0.069 and a coefficient of determination of 0.9127. A statistically significant superiority of this architecture over isolated convolutional or recurrent networks has been established. The effectiveness of the LSTM architecture for predicting anomalous events based on multidimensional data with an accuracy of 0.92 has been proven. The prospects of using multimodal and graph neural network models to increase the reliability of forecasts by taking into account spatial and temporal relationships are revealed. Practical significance. The results of the work form the basis for the creation of high-tech information and analytical systems of predictive analytics, applicable in various subject areas where the analysis of visual information with a timestamp is required. Discussion. The developed approaches demonstrate the potential for creating intelligent anomaly detection and decision support systems. The development of methods that are resistant to a limited amount of training data and the creation of approaches to quantify forecast uncertainty in complex dynamic environments are identified as promising areas.