Machine Learning-Based Conflict Detection And Trajectory Prediction For Aerial Vehicles

In recent years,general aviation has been promoted by Small Aircraft Transportation System(SATS)to broaden air transportation services.Under such circumstances,the National Airspace System(NAS)has undergone tremendous changes.In the past decade,Unmanned Aerial Vehicle(UAV)has rapidly developed into civilian applications,leading to more congested airspace at low and medium altitudes.Hence,aerial vehicles’ security issues pose a major challenge to the Air-Ground Integrated Vehicle Networks(AGIVN).Traditional surveillance technologies have obvious technical limitations in meeting requirements of future dense air traffic management(ATM).Therefore,an advanced automatic dependent surveillance-broadcast(ADS-B)technique has been proposed to track and monitor aerial vehicles in a more effective manner.This thesis proposes a grouping-based conflict detection algorithm based on the preprocessed ADS-B dataset,which realizes effective monitoring of aerial vehicles,and combines machine learning and deep learning algorithms to predict the trajectories of aerial vehicles,and then combines the predicted trajectory with the conflict detection algorithm to realize conflict prediction and greatly improve the flight safety of aerial vehicles.The specific research work is as follows:Firstly,because of the lack of a complete and extensive public flight dataset in China,this thesis assumes that the original ADS-B data collected by the self-built ADS-B integrated monitoring ground station is generally accurate.Through data preprocessing,a dataset suitable for this thesis is established.Secondly,a conflict detection algorithm based on time axis mapping is proposed in this thesis.The aerial vehicles conflict problem is transformed into the conflict problem of three-dimensional coordinates(X,Y,Z)in the space coordinate system,and the conflict regions on the three-dimensional coordinate axis are uniformly mapped.On the time axis,conflict time periods on X,Y,and Z axis are obtained,respectively.Then it is judged whether there is an intersection between conflict time periods on X,Y,and Z axis to determine whether there is a route conflict.However,considering that if two aerial vehicles are far apart,the probability of conflict between the two aerial vehicles is very low.In this scenario,it is meaningless to perform conflict detection on the two aerial vehicles and increase the time cost.Therefore,this thesis proposes the concept of grouping,which divides the space where the entire aerial vehicle group is divided into multiple areas,so that the large aerial vehicle group is divided into multiple small aerial vehicle groups.Considering that aerial vehicles at the edge of each group may also be close to each other,resulting in conflicts,aerial vehicles at the edge of each group are additionally divided into a group,and the grouping is finally completed.So far,this thesis combines the grouping idea with the conflict detection algorithm,and proposes a conflict detection algorithm based on grouping.The simulation results show that the average conflict probability meets the ATC standard,indicating that the grouping-based conflict detection algorithm proposed in this thesis can monitor the flight status of aerial vehicles in real time and effectively.Thirdly,in order to improve the security of the proposed algorithm,this thesis proposes trajectory prediction models based on CNN,LSTM,CNN-LSTM and LS algorithms,designs short-term and long-term trajectory prediction tasks,and trains and tests on the civil aviation dataset.The simulation results show that the CNN-LSTM-based prediction error is generally smaller than the prediction error based on CNN,LSTM and LS,and the prediction curve is more stable and more suitable for the actual trajectory.However,to reach the convergence point,the CNN-LSTM-based model takes more training time(6500 iterations),while the CNN and LSTM-based models require1000 and 6000 iterations,respectively,and the LS-based model even only takes a few seconds.It can be seen that although the CNN-LSTM-based predictor achieves higher accuracy,it trades for improved accuracy at the cost of time.The short-term prediction simulation results show that the four algorithms can accurately predict the trajectories.On the other hand,the error of long-term prediction increases with increasing lag time compared to short-term forecasts.In addition,in the long-term prediction task,when the lag time increases,the LSTM-based prediction error is the smallest,which indicates that the LSTM unit has a certain effectiveness in dealing with long-term prediction tasks.It is because CNN can extract the spatial characteristics of features and LSTM can extract temporal characteristics of features that the CNN-LSTM-based predictor can still maintain high prediction performance when the lag time increases.Finally,conflict prediction is performed using the CNN-LSTM-based trajectories.The simulation results show that the average conflict probability based on the predicted trajectory is lower compared to the one obtained using the actual trajectory.It can be inferred that the predicted trajectory obtained may be more ideal,while the actual trajectory will have sudden fluctuations.When using the predicted trajectory,its RMSE score is 0.893.In fact,there may be errors in the trajectory information obtained by the ADS-B receiver,but the error is usually about 100 meters,which has little impact on the trajectory prediction task.If the transmission process of the error is further considered,the RMSE score can increase from 0.893 to about 1 at most,which is still within the allowable error range of the conflict detection method.This shows that the predicted number of conflicts per second differs from the actual number of conflicts by less than one aerial vehicle(0.893<1).Hence,the conflict prediction method based on the predicted trajectory designed in this thesis can detect whether there is a conflict in the future time.