Risk Assessment Procedure of Final Approach to Landing Using Deep Learning

Abstract

Flight safety is the highest priority of all involved in aviation. All civil aviation authorities around the world have the responsibility to guarantee flight safety, but the flight data are considered confidential and regarded as trade secrets by airlines. Therefore, the publicly available sources are employed in the study to investigate the risk and related factors of flight safety during the approach to landing. Automatic Dependent Surveillance–Broadcast flight trajectory data and the Meteorological Terminal Aviation Routine Weather Report are used to analyze the risk associated with the final approach to landing. This study presents a procedure to conduct the data mining and risk assessment of flights landing at Taipei Songshan Airport (TSA). A deep learning method is adopted to tackle the problems associated with nonlinear regression that cannot be addressed by traditional statistical regression methods. The trajectory and glide path angle deviations of each flight are determined and applied to the model training. These two deviation variables are regarded as the evaluation standards, while the weather conditions are used as risk factors. Each weather factor in the deep learning models is weighted by the permutation feature importance analysis to capture how much influence each factor has on predictions from the model. The results show that the estimation accuracy of trajectory deviation with the 10-input model and glide path angle deviation with the 11-input model is over 90 and 95%, respectively. The trajectory deviations of runway 10 have higher impacts from temperature, precipitations, and crosswind.