Multiple machine learning modeling on near mid-air collisions: An approach towards probabilistic reasoning

This work presents a mathematical model to predict and explain Near Mid-Air Collisions (NMACs) based on the NASA Aviation Safety Reporting System (ASRS) database. The ASRS database contains more than 200,000 aviation incidents, which are used to learn how the combination of risk influencing factors (RIFs), such as crew size and component fatigue, affects the safety of airspace operations. Bayesian Networks (BNs) combine theory of probabilities with theory of graphs and are considered one of the most effective theoretical models in the fields of knowledge representation and reasoning with uncertainty. The resulting model allows to calculate the posterior probabilities of some targeted outputs, therefore providing a mathematically consistent framework to quantify and to compute with uncertainty the likelihood of incident occurrence over time when some factors are known. Furthermore, the bidirectional reasoning technology of BNs can calculate the posterior probabilities of its variables under the system incident condition, and find out the most likely combination that caused a NMAC. Finally, the resulting probabilistic models are compared with sixteen Machine Learning Algorithms, and advantageous properties were critically evaluated, such as a white-box reasoning and probability as a measure of certainty about the state of unobserved variables.