Exploring Pilot Workload Scenarios via Eye-Tracking: An Attempt at Inducing and Identifying Attentional Tunneling in the Cockpit

Kurzfassung

Given the profound impact of human errors and the essential role of operators in safety-critical domains, ensuring that operators are in a condition that allows them to adequately perform their tasks is a vital precaution. The timely identification of hazardous cognitive states can reduce accidents and enhance safety across various fields, including aviation. As workload and attentional tunneling are among the cognitive states most frequently associated with human error accidents in aviation, the purpose of this thesis is to explore the possibility of detecting these states using eye-tracking metrics. Attentional tunneling, a term commonly referenced in accident reports, is characterized by the excessive focus on a source of information, hypothesis, or goal to the disregard of other factors. Although previous research has demonstrated the recognition of workload in cockpit settings using eye-tracking metrics, attentional tunneling in simulator environments has rarely been explored. With this study, our aim was to propose and analyze scenarios for inducing and detecting attentional tunneling in simulator environments and to investigate the efficiency of transition frequency, mean saccade length, and entropy as a set of eye-tracking metrics for classifying workload and tunneling states. As tunneling triggering parameters, the proposed experiment design incorporated a workloadinducing secondary task and an ego-threatening factor in the form of negative auditory feedback on a focus task. Consequently, the occurrence of attentional tunneling was determined based on participants’ ability to notice visual cues related to abnormal cockpit behavior. This experimental framework was tested by 15 expert pilots, with data from 12 participants included in the eye-tracking and attentional tunneling analysis. Findings from the workload self-assessment measurements indicated the successful manipulation of workload between conditions. Moreover, the occurrence of attentional tunneling could be observed across one-third of the runs, suggesting that the proposed scenarios have proven efficient. The statistical analysis of the eye-tracking measurements revealed a significant decrease in the transition frequency and mean saccade length during high workload conditions. The occurrence of attentional tunneling, however, did not seem to significantly impact the recorded gaze measurements. Using the eye-tracking data, three machine-learning pipelines, including Support Vector Machines, Logistic Regression, and Bernoulli Naive Bayes, were trained and tested on their performance across two different classification problems: differentiating between low and high workload states and recognizing instances of attentional tunneling. With mean scores of approximately 50% for both accuracy and precision across all machine-learning approaches, the outcomes of the workload classification did not reach satisfactory performance. Similarly, the effectiveness of the logistic regression and SVM pipelines in classifying tunneling states showcased suboptimal results and a strong bias with relatively high accuracy mean scores and exceptionally low precision scores. Nevertheless, compared to the other two algorithms, the Bernoulli Naive Bayes demonstrated promising results that can be further investigated in future studies focusing on tunneling classification. Although the employed pipelines were unable to effectively classify the different cognitive states, the lessons learned have been instrumental in developing a strategy for subsequent improvements to our approach, mainly focused on data exploration and restructuring.