THE UAV RISK PREDICTION METHOD BASED ON TIME WINDOW EXTRACTION AND FLIGHT DATA GENERATION

This paper proposes a Generative Adversarial Network (GAN) method based on data sequence features for feature extraction from temporal windows and data generation for the prediction of UAV flight risks. This approach addresses the challenge of obtaining substantial flight data within time constraint and lack of similar type of data for reference during the development process of new types of UAV. By employing a time window technique to extract features from flight and utilizing GAN for data generation, this study overcomes the scarcity of flight, particularly the data related to risky incidents. The experimental results based on a real dataset demonstrate that the data generated by the proposed method, when input into an XGBoost classifier for training, can effectively predict the type and timing of risk incidents. Compared to data obtained by traditional GANs based on single time-point data, the generated data shows a distinct advantage. In terms of risk type prediction, the method achieved a F1 score of 0.991, and for the prediction of the timing of risk incidents, it achieved a F1 score of 0.803.