Runway safety improvements through a data driven approach for risk flight prediction and simulation
MetadataShow full item record
Runway overrun is one of the most frequently occurring flight accident types threatening the safety of aviation. Sensors have been improved with recent technological advancements and allow data collection during flights. The recorded data helps to better identify the characteristics of runway overruns. The improved technological capabilities and the growing air traffic led to increased momentum for reducing flight risk using artificial intelligence. Discussions on incorporating artificial intelligence to enhance flight safety are timely and critical. Using artificial intelligence, we may be able to develop the tools we need to better identify runway overrun risk and increase awareness of runway overruns. This work seeks to increase attitude, skill, and knowledge (ASK) of runway overrun risks by predicting the flight states near touchdown and simulating the flight exposed to runway overrun precursors. To achieve this, the methodology develops a prediction model and a simulation model. During the flight training process, the prediction model is used in flight to identify potential risks and the simulation model is used post-flight to review the flight behavior. The prediction model identifies potential risks by predicting flight parameters that best characterize the landing performance during the final approach phase. The predicted flight parameters are used to alert the pilots for any runway overrun precursors that may pose a threat. The predictions and alerts are made when thresholds of various flight parameters are exceeded. The flight simulation model simulates the final approach trajectory with an emphasis on capturing the effect wind has on the aircraft. The focus is on the wind since the wind is a relatively significant factor during the final approach; typically, the aircraft is stabilized during the final approach. The flight simulation is used to quickly assess the differences between fight patterns that have triggered overrun precursors and normal flights with no abnormalities. The differences are crucial in learning how to mitigate adverse flight conditions. Both of the models are created with neural network models. The main challenges of developing a neural network model are the unique assignment of each model design space and the size of a model design space. A model design space is unique to each problem and cannot accommodate multiple problems. A model design space can also be significantly large depending on the depth of the model. Therefore, a hyperparameter optimization algorithm is investigated and used to design the data and model structures to best characterize the aircraft behavior during the final approach. A series of experiments are performed to observe how the model accuracy change with different data pre-processing methods for the prediction model and different neural network models for the simulation model. The data pre-processing methods include indexing the data by different frequencies, by different window sizes, and data clustering. The neural network models include simple Recurrent Neural Networks, Gated Recurrent Units, Long Short Term Memory, and Neural Network Autoregressive with Exogenous Input. Another series of experiments are performed to evaluate the robustness of these models to adverse wind and flare. This is because different wind conditions and flares represent controls that the models need to map to the predicted flight states. The most robust models are then used to identify significant features for the prediction model and the feasible control space for the simulation model. The outcomes of the most robust models are also mapped to the required landing distance metric so that the results of the prediction and simulation are easily read. Then, the methodology is demonstrated with a sample flight exposed to an overrun precursor, and high approach speed, to show how the models can potentially increase attitude, skill, and knowledge of runway overrun risk. The main contribution of this work is on evaluating the accuracy and robustness of prediction and simulation models trained using Flight Operational Quality Assurance (FOQA) data. Unlike many studies that focused on optimizing the model structures to create the two models, this work optimized both data and model structures to ensure that the data well capture the dynamics of the aircraft it represents. To achieve this, this work introduced a hybrid genetic algorithm that combines the benefits of conventional and quantum-inspired genetic algorithms to quickly converge to an optimal configuration while exploring the design space. With the optimized model, this work identified the data features, from the final approach, with a higher contribution to predicting airspeed, vertical speed, and pitch angle near touchdown. The top contributing features are altitude, angle of attack, core rpm, and air speeds. For both the prediction and the simulation models, this study goes through the impact of various data preprocessing methods on the accuracy of the two models. The results may help future studies identify the right data preprocessing methods for their work. Another contribution from this work is on evaluating how flight control and wind affect both the prediction and the simulation models. This is achieved by mapping the model accuracy at various levels of control surface deflection, wind speeds, and wind direction change. The results saw fairly consistent prediction and simulation accuracy at different levels of control surface deflection and wind conditions. This showed that the neural network-based models are effective in creating robust prediction and simulation models of aircraft during the final approach. The results also showed that data frequency has a significant impact on the prediction and simulation accuracy so it is important to have sufficient data to train the models in the condition that the models will be used. The final contribution of this work is on demonstrating how the prediction and the simulation models can be used to increase awareness of runway overrun.