Adaptive Training for Aircraft Simulation: A Data-Driven Approach for Pilot Workload Assessment

Adaptive training systems have the potential to revolutionize pilot training by dynamically adjusting simulation difficulty based on real-time assessments of individual needs and performance. This study investigates the relationship between physiological signals and varying difficulty levels in flight simulations, exploring the potential to assess workload. A publicly available dataset containing multimodal signals (electrocardiography, electrodermal activity, photoplethysmography, accelerometry, electromyography, oculometry, respiration, and flight data) was analyzed. The data originated from 35 pilots performing ILS approaches at four difficulty levels with progressively increasing wind, turbulence, and visibility restrictions. This study employed a rigorous data preparation and analysis pipeline. Preprocessing included time series validation and signal-specific filtering guided by established literature. Comprehensive feature extraction techniques were applied, encompassing statistical measures and specialized eye-tracking features. To find the most significant features and enhance interpretability, the feature selection method was utilized. Finally, various machine learning models were trained and evaluated to assess classification performance and the feasibility of accurately discriminating workload levels. Due to the inherent differences between simulated and real flight, interpreting physiological responses in a simulator context requires specific considerations; in fact, eye-tracking emerged as the most informative physiological signal for workload assessment. While biosignals alone resulted in moderate multiclass classification accuracy (64%), integrating flight data significantly boosted performance (96%), highlighting the value of combining objective performance metrics with physiological indicators. Binary classification (low vs. high difficulty) using biosignals alone yielded 91% accuracy, indicating clear distinctions between very different workloads while also highlighting the challenge of discerning more similar difficulty levels. Eye-tracking within VR simulators offers a non-invasive, cost-effective tool for analyzing a pilot's cognitive state. When coupled with flight data, this enables enhanced workload discrimination. These findings support the feasibility of AI-driven adaptive training systems that leverage eye-tracking and flight data, offering the promise of highly personalized and efficient pilot training. This research lays the groundwork for future systems that dynamically adjust to individual pilot needs, optimizing training outcomes and potentially reducing training costs.