Automating pilot function performance assesssment using fuzzy systems and a genetic algorithm

• Main Author: Zaspel, Joachim C.
• Other Authors: Funk, Kenneth H.
• Language: en_US
• Published: 2012
• Subjects: Airplanes > Control systems; Fuzzy systems; Genetic algorithms; Decision-making > Data processing; Decision support systems
• Online Access: http://hdl.handle.net/1957/33694

Modern civil commercial transport aircraft provide the means for the safest of all forms of transportation. While advanced computer technology ranging from flight management computers to warning and alerting devices contributed to flight safety significantly, it is undisputed that the flightcrew represents the most frequent primary cause factor in airline accidents. From a system perspective, machine actors such as the autopilot and human actors (the flightcrew) try to achieve goals (desired states of the aircraft). The set of activities to achieve a goal is called a function. In modern flightdecks both machine actors and human actors perform functions. Recent accident studies suggest that deficiencies in the flightcrew's ability to monitor how well either machines or themselves perform a function are a factor in many accidents and incidents. As humans are inherently bad monitors, this study proposes a method to automatically assess the status of a function in order to increase flight safety as part of an intelligent pilot aid, called the AgendaManager. The method was implemented for the capture altitude function: seeking to attain and maintain a target altitude. Fuzzy systems were used to compute outputs indicating how well the capture altitude function was performed from inputs describing the state of the aircraft. In order to conform to human expert assessments, the fuzzy systems were trained using a genetic algorithm (GA) whose objective was to minimize the discrepancy between system outputs and human expert assessments based on 72 scenarios. The resulting systems were validated by analyzing how well they conformed to new data drawn from another 32 scenarios. The results of the study indicated that even though the training procedure facilitated by the GA was able to improve conformance to human expert assessments, overall the systems performed too poorly to be deployed in a real environment. Nevertheless, experience and insights gained from the study will be valuable in the development of future automated systems to perform function assessment. === Graduation date: 1998