Research On Early Warning Of Airliner Runway Overrun Events Based On QAR Data And Deep Learning

With the rapid development of the civil aviation industry,more and more attention is paid to civil aviation flight safety.The runway overrun is one of the most typical safety accidents during the landing phase,and its early warning is of great significance to the safety of civil aviation.Quick Access Recorder(QAR),as an airborne device,records detailed data during the flight of an aircraft,and has been widely used in recent years to conduct flight safety-related research.However,the existing runway overrun early warning researches often highly rely on the prior knowledge of flight experts and focus on retrospective analysis.The prediction accuracy is unsatisfactory and it is difficult to apply in the real-time flight process of civil aviation.In response to the above problems,this thesis proposes proactive methods for aircraft landing speed and landing distance prediction,so as to enable early warning of runway overrun.The main content of this thesis is as follows:To deal with the problems of QAR data complexity,unguaranteed integrity,and complicated spatio-temporal inter-dependencies,the data is preprocessed through cleaning,transformation,and feature selection.Landing speed is one of the main indicators for early warning of runway overrun safety incidents.For the runway overrun problem of caused by excessive speed,this thesis proposes an aircraft landing speed sequence prediction model GLC(GBDT LSTM codec)based on QAR data and a deep sequence-to-sequence method.GLC uses GBDT to select features that are important to landing speed,and uses tree node splitting information gain to calculate feature global importance.The selected features are normalized and imported into the deep LSTM encoding and decoding sequence model.The encoder reads the input sequence and encodes it into a fixed-length context vector,while the decoder decodes the vector and outputs the predicted sequence.On the basis of the GLC model,this thesis further proposes an aircraft landing distance prediction model LSTRO(Layered Sequence Spatio-Temporal for Runway Overrun)that integrates multi-layer spatio-temporal codec and TG-Attention.LSTRO considers the spatio-temporal dependence between CNN-LSTM perception features,and uses TG-Attention to focus on pilots’ operations and flight trends,and at the same time builds a layered codec to integrate information and feedback,thereby greatly improving prediction performance.The model proposed in this thesis is evaluated on the A321 aircraft data set,and the superiority and effectiveness of the model are verified through stripping experiments and comparative experiments.The innovations of the thesis are summarized as follows:(1)In view of the problem of excessive QAR parameters and excessive reliance on flight experts for feature selection,the ensemble learning method GDBT is used to select important features to achieve preliminary automatic selection of features and reduce noise interference caused by excessive QAR parameters.(2)Considering the fact that existing research methods such as RNN focus on QAR temporal features and are not sensitive to the spatial information of different regions at the same time,this thesis fuses CNN and LSTM to extract local spatial features while retaining the temporal information of QAR data as much as possible.At the same time,an innovative layered codec model is proposed,where the upper codec integrates the information of the lower codec and then feeds it back to improve the accuracy of the final prediction sequence.(3)This thesis proposes a time-series attention module TG-Attention that integrates improved self-attention mechanism and global attention mechanism.Through the organic fusion of two attention mechanisms,this module can capture both the influence of the pilot’s flight operations at every moment and overall flight trends on future prediction sequences.