Machine Learning based Prediction of Situational Awareness in Pilots using ECG Signals

Reducing aviation fatalities requires a high level of reliable real-time monitoring so that events can be predicted and prevented before they can occur. Situational awareness is essential in the cockpit where manual and autonomous operations co-exist. Many interventions and countermeasures have been designed into cockpits to enhance pilot awareness and performance. This study aims to analyse pilot and copilot teams' awareness by using physiological data which was collected in a flight simulator to train models to predict when pilots are in a state of Channelised Attention (CA), Diverted Attention (DA), and Startle/Surprise (SS). Electrocardiogram (ECG) signals collected for 18 subjects were processed in preparation to develop a comprehensive tool which utilises active Line Oriented Flight Training (LOFT) data to evaluate machine learning tools which are capable of predicting pilot awareness response. A combination of linear, non-linear, binary and multi-class classification were applied to this data. The results indicate that while all classifiers produced stable results, Decision Tree(DT) far outperformed the others. Further analyses revealed that the maximum value for ECG was the most important feature used by all classifiers evaluated for importance in training the classification models. However, for DT which was the best performing classifier both maximum and minimum ECG values were the most important features in predictions made by this model.