Date of Award

Fall 2021

Access Type

Dissertation - Open Access

Degree Name

Doctor of Philosophy in Aviation

Department

College of Aviation

Committee Chair

Dothang Truong, Ph.D.

First Committee Member

David Esser, Ph.D.

Second Committee Member

Mark A. Friend, Ed.D.

Third Committee Member

Robert W. Maxson, Ph.D.

Abstract

Pilots, flight attendants, and passengers can be exposed to toxic compounds when the bleed air that supplies the cabin and flight deck is contaminated with pyrolyzed hydraulic fluid or oil from turbine jet engines. These fume events occur sporadically and can result in acute or chronic exposure in air crews and can have catastrophic consequences if flight crew members become impaired or incapacitated. The purpose of this research was to explore unstructured textual data and identify important factors associated with these events. Models using machine learning algorithms were developed and tested using variables gleaned from the text mining process and variables found in self-reported aviation incidents.

Safety reports from flight and cabin crews working in 14 C.F.R. § 121 Domestic, Flag, and Supplemental Operations during 2015-2019 were downloaded from the Aviation Safety Reporting System (ASRS). Narratives from these reports were explored using the text mining process in SAS®Text Miner to identify potentially new factors associated with the occurrence of fume events. The text mining process included text parsing, text filtering, text clustering, and text topic. The identified factors were combined with variables from the ASRS reports to develop six models. These models used decision tree, gradient boosting, logistic regression, and random forest algorithms.

Values for misclassification rate, receiver operating characteristic curves, and lift curves were used to assess model accuracy and predictive power to determine the best-performing model. Four models produced similar results with accuracy above 96 percent. The top four performing models were gradient boosting, random forest, logistic regression, and a 7-branch decision tree model.

Sensory perception was found to be the most important factor in all four top-ranking models for the occurrence of fume events. The cabin affected and power change factors were also listed in the top ten factors in four of the models with varying degrees of importance. Four other factors, including aircraft action, passenger disruption, system anomaly, and engine issue, were associated with the occurrence of fume events in three of the top four models. The identification of sensory perception, power change, aircraft action, and engine issue is consistent with previous research in fume events. This research identified three new factors associated with the occurrence of fume events: cabin affected, passenger disruption, and system anomaly. These factors can be used in the identification of fume events and to educate and increase awareness of potential events for flight and cabin crew members. They can also be included as variable fields in a national database to capture information regarding the occurrence of these events. Each of these activities could contribute to the health and safety of crews and passengers, reduce flight disruptions due to fume events, and limit financial losses due to flight disruptions.

Download

Share

COinS