Feature Selection for Predicting Pilot Mental Workload

摘要

Advances in technology have the cockpits of the aircraft in the Air Force inventory increasingly complex. Consequently, mental demands on the pilot have risen. In some cases, mental demands were so overwhelming that pilots have forgotten basic flying techniques, such as G-straining maneuvers. The results have been fatal. Recent research in this area has involved collecting psychophysiological features, such as electroencephalography (EEG), heart, eye and respiration measures, in an attempt to identify pilot mental workload. This thesis focuses on feature selection and reduction of the psycophysiological features and subsequent classification of pilot mental workload on multiple subjects over multiple days. A stepwise statistical technique and the signal-to- noise ratio (SNR) saliency metric were used to reduce the number of features required for classification. Factor analysis was used to compare the variables chosen by the discriminant procedure and the SNR metric as applied to a neural network. A total of 151 psychophysiological features were derived from data collected during an actual flight study. The original flight study contained three workload levels, low, medium and high. These levels were aggregated into two categories of pilot mental workload, low/medium and high. Mental workload associated with each flight segment was determined by difficult of task.