Gryshmanov Y. The method of automated forecasting of adverse aviation events to improve flight safety in air traffic control

The dissertation is aimed at solving the actual scientific problem of improving the methods and models of automated forecasting of unfavorable aviation events in flight based on methods and models of in-depth training to improve the quality of risk assessment for flight safety. For the first time, the method of forecasting unfavorable aviation events in flight based on convolutional and recurrent neural networks has been developed. Unlike the well-known, the classification of unfavorable aviation events is based on the use of a convolutional neural network, and for the initial setup of the vector layer of the hybrid model of forecasting, a pre-trained layer of recurrent neural network. The method for forming a training sample for training the deep hybrid neural network for forecasting unfavorable aviation events in flight, which, unlike the known ones, provides the construction of a dictionary of text messages about unfavorable aviation events, using the measure of the importance of words and the vector model of text messages about adverse aviation events on the labeled a data set using a vector representation of words. The metod of automated forecasting of unfavorable aviation events in the flight, which, unlike the known ones, is based on the knowledge-oriented representation of the stages of risk assessment for flight safety, has been further developed. This enables intelligent data processing to improve the accuracy and completeness of the automated classification of adverse aviation events in flight. The combination of advanced methods is the scientific essence of the formalization of the processes of automated forecasting of adverse aviation events in the flight. Using the results of the study will improve the quality of risk assessment for flight safety through the introduction of automated prediction of unfavorable aviation events in flight in automated air traffic control systems. The results of the performed calculations and simulation modeling of the evaluation of the effectiveness of the methods and models, as well as the practical implementation of the results, confirmed the adequacy of the proposed methods and models of automated prediction of adverse aviation events in the flight based on the methods and models of in-depth training. Realization of the developed method of forecasting unfavorable aviation events in flight on the basis of convolutional and recurrent neural networks allowed to automate the process of forecasting unfavorable aviation events. The implementation of the advanced method for forming a training sample for training a deep hybrid neural network for forecasting adverse aviation events in flight allowed the use of the vector model of the training sample as the basis for studying the deep hybrid neural network for forecasting unfavorable aviation events. Implementation of metod for automated forecasting of adverse aviation events in flight in the subsystem of risk assessment for flight safety has allowed gain in accuracy and completeness of the classification of unfavorable aviation events on average up to 5%.