Monitoring pilots' mental workload in real flight conditions using multinomial logistic regression with a ridge estimator.

Abstract

Piloting an aircraft is a cognitive task that requires continuous verbal, visual, and auditory attentions (e.g., Air Traffic Control Communication). An increase or decrease in mental workload from a specific level can alter auditory and visual attention, resulting in pilot errors. The objective of this research is to monitor pilots' mental workload using advanced machine learning techniques to achieve improved accuracy compared to previous studies. Electroencephalogram (EEG) data were recorded from 22 pilots operating under visual flight rules (VFR) conditions using a six dry-electrode Enobio Neuroelectrics system, and the Riemannian artifact subspace reconstruction (rASR) filter was used for data cleaning. An information gain (IG) attribute evaluator was used to select 25 optimal features out of 72 power spectral and statistical extracted features. In this study, 15 classifiers were used for classification. Multinomial logistic regression with a ridge estimator was selected, achieving a significant mean accuracy of 84.6% on the dataset from 17 subjects. Data were initially collected from 22 subjects, but 5 were excluded due to data synchronization issues. This work has several limitations, such as the author did not counter balance the order of scenario, could not control all the variables such as wind conditions, and workload was not stationary in each leg of the flight pattern. This study demonstrates that multinomial logistic regression with a ridge estimator shows significant classification accuracy (p < 0.05) and effectively detects pilot mental workload in real flight scenarios.