Breaking Down the Barriers To Operator Workload Estimation: Advancing Algorithmic Handling of Temporal Non-Stationarity and Cross-Participant Differences for EEG Analysis Using Deep Learning
Date of Award
Doctor of Philosophy (PhD)
Department of Electrical and Computer Engineering
Brett Borghetti, PhD
This research focuses on two barriers to using EEG data for workload assessment: day-to-day variability, and cross- participant applicability. Several signal processing techniques and deep learning approaches are evaluated in multi-task environments. These methods account for temporal, spatial, and frequential data dependencies. Variance of frequency- domain power distributions for cross-day workload classification is statistically significant. Skewness and kurtosis are not significant in an environment absent workload transitions, but are salient with transitions present. LSTMs improve day- to-day feature stationarity, decreasing error by 59% compared to previous best results. A multi-path convolutional recurrent model using bi-directional, residual recurrent layers significantly increases predictive accuracy and decreases cross-participant variance. Deep learning regression approaches are applied to a multi-task environment with workload transitions. Accounting for temporal dependence significantly reduces error and increases correlation compared to baselines. Visualization techniques for LSTM feature saliency are developed to understand EEG analysis model biases.
DTIC Accession Number
Hefron, Ryan G., "Breaking Down the Barriers To Operator Workload Estimation: Advancing Algorithmic Handling of Temporal Non-Stationarity and Cross-Participant Differences for EEG Analysis Using Deep Learning" (2018). Theses and Dissertations. 1949.