PGTNet: a physiology-guided parallel Graph-Transformer network for multimodal fatigue recognition of aviation maintenance engineers

Fatigue severely degrades operational reliability in high-risk operational settings such as aviation maintenance. Although multimodal physiological signals support objective fatigue assessment, many existing models emphasize temporal patterns without explicitly modeling dependencies among physiological features across modalities. We propose a Physiology-Guided Parallel Graph–Transformer Network (PGTNet) for multimodal fatigue recognition, which incorporates physiologically motivated fatigue-related features derived from multimodal physiological signals. Parallel DGCN and Transformer branches jointly model physiological feature relationships and long-range temporal dependencies, and their representations are integrated through a bidirectional cross-attention mechanism. Under the predefined subject-independent partition, PGTNet achieves an Accuracy of 0.8778 and a Macro-AUC of 0.9648. Eight-fold subject-wise cross-validation yields an average Accuracy of 0.8592 ± 0.0112 across the eight held-out subjects. These results support the effectiveness of PGTNet within the present dataset.