Prediction Of Aircraft Hard Landing Based On Mechine Learning

Hard landing,a common flight accident during the landing phase,refers to a large landing vertical acceleration.The existing hard landing detection standards are post-event safety measures and cannot achieve pre-prevention.Therefore it is necessary to build a prediction model of the hard landing,which based on the existing large number of flight data,realize the pre-prevention of hard landing.First,reviewed the literature on aircraft hard landing prediction and conducted in-depth research on aircraft hard landing from many aspects.Based on the airplane flying handbook,the characteristics of the pilot’s operation during the approach and landing were analyzed;The landing force analysis of aircraft under the two conditions of symmetric landing and asymmetric landing were established respectively;According to relevant literature,the influencing factors and main flight parameters of hard landing were summarized,and three types of flight parameters closely related to hard landing were selected as data features;Finally,the defects of the existing hard landing judgment standards were analyzed.Second,according to the selected flight parameters,the B737 QAR raw data was filtered and preprocessed.Firstly,the QAR data for approach and landing were filtered according to the radio altitude,and box-plot and linear interpolation were used to eliminate vacancies and outliers to avoid erroneous data in the sample;then,the individual flight parameters were nondimensionalization to eliminate the influence of different dimensions and units of parameters;finally,principal component analysis was used to extract the key features of the data as the samples for the prediction model.Third,the hard landing prediction models of the approach and landing phases were built respectively.The feature matrices after data processing were randomly divided into 5 training sets and 1 test set,cross-validation was carried out to prevent the model from over-fitting to the learning data.The isolation forest method was implemented through Python 3.7,and the prediction model of low-altitude approach phase and the prediction model of roundout and landing phase were obtained by learning the training set.Finally,the model was applied based on the test set.Firstly,ROC curve was used to preliminarily evaluate the model performance,and the model with the best performance was selected.Combined with the flight operation quality assurance,the prediction of ultra limits incidents in the approach phase and the prediction of hard landing were analyzed,besides the prediction accuracy and pre-warning time of each ultra limits project were obtained;based on the specific prediction results of ultra limits events,corresponding safety recommendations were put forward;Finally,through comparative experiments,the superiority of the isolation forest model was verified.The model realizes the prediction of unsafe events such as hard landing during the landing phase of aircrafts.In the future,in-depth research can be conducted in terms of crew operation data,real-time early warning systems,and pilot characteristics.