Flight Data Driven System Identification Using Neural Networks for Landing Safety Assessment

Abstract

Multiple studies and reports have identified threshold exceedance precursors that significantly contribute to flight accidents. Reducing the occurrence of these precursors may potentially improve flight safety. This may be achieved by increasing pilots exposure to them. This in turn can be realized by training pilots with flights that have multiple accident precursors. To that end, this work focuses on the development of simulated flights that could eventually be integrated within a simulation environment to help increase pilots’ awareness about the precursors that lead to runway excursion events. Simulating flights with precursors require aircraft dynamic behavior to accurately model real world flying conditions. This requires model inputs to be mapped to the states based on real flight data. Real world flying conditions are represented by the flight operational quality assurance (FOQA) data. For this study, FOQA data is used to train long short term neural network model (LSTM) to simulate the aircraft behavior in final approach. Final approach is select as the flight phase of interest based on its significance to most frequent flight accidents such as runway excursion. A neural network model is used to simulate the final approach by sequentially predicting future aircraft states based on current aircraft states and full control trajectories. The accuracy of the aircraft model is evaluated based on the accuracy of predicted and actual state trajectories. Then the aircraft model is used to compare number of precursors triggered by simulated and actual flights. This comparison is used to evaluate the potential for simulated flights to be used for identifying accident precursors and increasing awareness to the precursors.