Research On Deep Learning Based Flight Training Maneuver Identification Method

According to the strategic goal of developing a strong civil aviation country,the first two decades of this century are an important period for the development of China’s civil aviation,bringing unprecedented development opportunities to China’s civil aviation industry,which has given rise to the expansion of the number of domestic flight training schools as well as the scale of training.The flight data generated by flight training can fully reflect the flight status of the aircraft.How to identify flight training actions from the massive amount of flight data,assist flight training and further improve flight training quality is a common research topic for flight training schools.With the rise of artificial intelligence technology,the method of deep learning is able to rely on the powerful computing power of computers to deeply mine massive data for data classification and identification,providing a new technical direction for the identification of flight training maneuvers.First,this thesis collects real flight training data and uses Kalman filtering algorithm to remove the wild values and K-approximation algorithm to fill the missing values to improve the quality of the flight data.Using the principle of motion decomposition,we use the data labeling aid to label the flight training maneuvers from both horizontal and vertical branches to build a highly usable flight training maneuver dataset.Secondly,by studying the convolutional neural network model in the deep learning model,the flight training maneuver recognition method based on two-branch convolutional neural network is proposed,which divides the multiple flight parameters required to recognize complex maneuvers into horizontal and vertical parts,and carries out the recognition of basic flight training maneuvers from these two branches separately,and then synthesizes the recognition results of the two branches to finally obtain the complex flight maneuvers of the aircraft.The model is trained using the flight training maneuvers dataset constructed in this thesis,and the key parameters of the model are optimized according to the recognition accuracy,so as to obtain the optimal model parameters.The experiments show that this thesis conducts training and recognition from two branches,which speeds up the computation,reduces the model training time and has a high recognition accuracy.Finally,in order to check the correctness of the algorithm,the results of the algorithm are visualized using the existing development platform to develop a flight training maneuver visualization system,and the development process is explained in detail.