Pilot Workload Classification Using Artificial Neural Networks in a Simulated Scud Hunt Mission

Real-time knowledge of an operator’s mental workload state could increase the efficiency and efficacy of pilots in advanced fighter aircraft by controlling the nature and type of information made available to them in different mental workload states. Most importantly, it is critical to know if a pilot is operating in an acceptable mental workload range or in an overload condition. Earlier studies summarized data that suggested that a mental workload “redline” can be defined based on observing when critical performance measures start to enter an error-prone range. The present study evaluated whether a redline could be predicted from psychophysiological and behavioral aspects of pilot state. Approaches, such as Baysian estimation and linear statistical techniques, have been used for predicting cognitive workload, but they have not achieved the required accuracy needed to implement them. Neural networks have several advantages over these approaches. They are adaptive and have the ability to generalize, therefore they have the capability of capturing cognitive interactions occurring during complex dynamic tasks performed by fighter pilots. The present study used psychophysiological and performance measures to predict whether pilots were operating above or below a critical mental workload redline. The classification was performed using a backpropagation neural network. Psychophys-iological measures included multiple channels of fast Fourier transforms from continuous electroencephalograph recordings. Peripheral measures of heart rate, eye blink and respiration rate and performance measures such as stick variability also were used. The network had two output nodes, which denoted an acceptable or overload of cognitive workload. Fifteen college-aged participants performed a simulated combat task in a medium fidelity air-toground combat simulator. Their task consisted of flying to a series of waypoints displayed on a Tactical Situation Display and identifying and destroying a SCUD launcher among a group of three vehicles at each waypoint. The pilot’s workload and critical performance was manipulated by controlling the aircraft’s speed (215, 325, 380, 435, 490, and 600 knots) on each 162 s trial. Redline speed was determined individually for each pilot by an objective procedure based on the number of targets destroyed. Each pilot’s data were analyzed individually by a neural network. Following training, performance was evaluated on a separate independent portion of the input data. Each network’s ability to classify trials as above or below redline was assessed. Classification accuracy on transfer trials was very high. Mean classification was 98.6 %, ranging from 97.2 % to 99.4 % for individual pilots. Neural networks were able to successfully predict each pilot’s mental workload redline point. The input variables have the potential of being utilized in a variety of tasks, and neural nets have the potential for being a reliable method for controlling pilot interfaces and information flow if the networks perform similarly in other tasks.