Workload State Classification With Automation During Simulated Air Traffic Control

Abstract

Real-time operator workload assessment and state classification may be useful for decisions about when and how to dynamically apply automation to information processing functions in aviation systems. This research examined multiple cognitive workload measures, including secondary task performance and physiological (cardiac) measures, as inputs to a neural network for operator functional state classification during a simulated air traffic control (ATC) task. Twenty-five participants performed a low-fidelity simulation under manual control or 1 of 4 different forms of automation. Traffic volume was either low (3 aircraft) or high (7 aircraft). Participants also performed a secondary (gauge) monitoring task. Results demonstrated significant effects of traffic volume (workload) on aircraft clearances (p < .01) and trajectory conflicts (p < .01), secondary task performance (p < .01), and subjective ratings of task workload (p < .01). The form of ATC automation affected the number of aircraft collisions (p < .05), secondary task performance (p < .01), and heart rate (HR; p < .01). However, heart rate and heart rate variability measures were not sensitive to the traffic manipulation. Neural network models of controller workload (defined in terms of traffic volume) were developed using the secondary task performance and simple heart rate measure as inputs. The best workload classification accuracy using a genetic algorithm (across all forms of ATC automation) was 64%, comparable to prior work. Additional neural network models of workload for each mode of ATC automation revealed substantial variability in predictive accuracy, based on the characteristics of the automation. Secondary task performance was a highly sensitive indicator of ATC workload, whereas the heart rate measure appeared to operate as a more global indicator of workload. A limited range of cardiac response might be sufficient for the demands of the brain in ATC. The results have applicability to design of future adaptive systems integrating neural-network-based workload state classifiers for multiple forms of automation.