| dc.description.abstract | A commercial civil aviation flight typically goes through multiple phases from flight planning to the final landing, including push back, takeoff, climb, cruise, descent, final approach, and landing. Among them, the final approach and landing have been cited as the riskiest phases. Among the incidents during the approach and landing phases, hard landing and runway overrun are the two most frequent; each of them accounts for 11.1 % and 12.0 % of the incidents. Predictive hard landing and landing distance models can provide pilots with early warnings so that timely corrective actions can be taken to mitigate the adverse consequences. Due to the high workload and mental fatigue, pilots sometimes fail to get the correct sense of energy level during the approach, leading to high speed exceedance (HSE), which is often cited as the cause of unstable approaches, hard landing, and long landing. Identifying the precursors of the HSE events is of great importance to pilots and aviation safety organizations. Besides HSE, there are other factors that can lead to accidents, such as aircraft system failure due to component fatigue, inappropriate pilot operations (e.g, late flap setting), and adverse weather conditions (e.g, low invisibility). A careful analysis of the anomalous flights helps flight safety investigators to understand the causes of the accidents and further regulate pilot operations to avoid similar maloperations in the future.
Machine learning models have gained a lot of attention due to their superior performance in different applications. Thus in this work, machine learning models are constructed using flight data. The aim of this dissertation is to utilize the robustness of machine learning models to develop a rigorous and intelligent aviation safety assessment system, with model uncertainty quantified and model functioning mechanisms interpreted. The research effort in this dissertation is divided into four objectives. The first objective is to predict hard landing with a probabilistic Bayesian recurrent neural network model. The second objective is to predict landing distance with a multi-step recurrent neural network model. The third objective is to develop a probabilistic autoencoder model to identify the anomalous flights in the landing phase. The fourth objective is to identify the precursors of an HSE event, both in terms of time of occurrence and the associated parameters, using a combination of convolutional and recurrent neural networks and a graphical technique that helps to explain the model reasoning. The proposed models are compared with state-of-art models, and the numerical results demonstrate satisfactory performance. | |
