Effective mental workload recognition is essential for reducing human error and enabling dynamic function allocation in human–machine systems. In this study, a multimodal physiological sensor-based mental workload recognition approach was developed to discriminate among underload, moderate load, and overload using functional near-infrared spectroscopy (fNIRS), electrocardiography (ECG), and eye-movement signals. Twenty-seven flight trainees participated in a flight simulation tracking task. Initially, differences in physiological response patterns across different mental workload levels were identified through the analysis of physiological signals. Subsequently, repeated-measures ANOVA and multicollinearity analysis were used to extract critical features, and the machine learning algorithm was utilized to evaluate the classification performance of different feature combinations. Finally, an explainability assessment of the model was performed by employing feature contributions derived from the Shapley additive explanations (SHAPs) method. The main findings are as follows: 1) the prefrontal cortex (PFC) activation level is highest under overload conditions, followed by underload, and lowest under moderate load. Under overload conditions, the activation level of the right PFC is higher than the left. In contrast, under moderate load and underload conditions, the activation level of the left PFC is higher than the right. 2) The best performance was obtained by the model trained on three types of modal features: response amplitude ( β ), heart rate variability (HRV), and pupil diameter. The model achieved a recognition accuracy of 85.76% under 10-fold cross-validation and 78.57% under leave-one-subject-out (LOSO) cross-validation. 3) SHAP analysis clarified that the response amplitude ( β ) of CH18 and CH20 in the left PFC BA10 region, along with Mean NN, W, and average pupil diameter of both eyes, are the most effective features for mental workload recognition.