Crosswind landing is one of the important training subjects for flight trainees in the intermediate stage of flight training. Accurately assessing the workload of trainees during this training is essential for mitigating human operational risks and enhancing the overall quality and effectiveness of the training process.To address this requirement, a non-invasive method for predicting the workload of flight trainees was proposed, which integrated facial action units (AUs), gaze data, and head posture data. First, a flight training scenario for crosswind landings was constructed. Based on this scenario, the factors that influence workload within the “human-machine-environment” system were identified.Subsequently, a comprehensive evaluation system for the workload during crosswind landings was established, integrating both subjective and objective data. Building on this foundation, a flight experiment using a crosswind landing simulator was designed to collect objective data on trainees' AUs, gaze, and head posture, along with subjective workload assessments based on the NASA-TLX scale.The CRITIC model was then employed to determine the feature weights of the six dimensions in the NASA-TLX scale. According to the levels of workload, flight trainees were categorized into three groups: low, medium, and high. The symmetric dot pattern (SDP) was utilized to perform a nonlinear mapping of the objective data into the polar coordinate space, enabling effective data fusion.Furthermore, fuzzy C-means clustering analysis was applied to explore the correlation between the fused data and subjective workload, leading to the generation of a fused feature pattern that measures workload and its dimensions.Finally, based on this feature pattern, four machine learning algorithms, including support vector machine (SVM), random forest, logistic regression, and extreme gradient boosting, were employed to construct a workload prediction model for flight trainees during crosswind landing. The experimental results demonstrate that the workload prediction model based on the “SVM + fusion feature pattern” achieves the highest accuracy, effectively measuring the workload of trainees in crosswind landing flight training.