Physics-informed and transformer-based machine learning for real-time structural health monitoring of self-sensing cementitious composites

This study evaluates Transformer and Physics-Informed Neural Network (PINN) models for real-time structural health monitoring (SHM) of self-sensing cementitious composites with hybrid Silicon Carbide and Graphite. While the Transformer model achieved higher crack classification accuracy (96.64 %) and stress prediction performance (R² = 0.8727, RMSE = 0.0839 MPa), the PINN model produced physically consistent outputs with more stable and concentrated error distributions. Voltage drop emerged as the most sensitive feature for crack detection, but its behavior varied by damage stage—increasing during crack initiation, then decreasing during macrocrack formation due to electrical path disruption. Environmental parameters (temperature, humidity) also showed statistically significant influence (p < 0.0001), particularly during late-stage damage progression. Violin plots and time-series trends revealed that strain loses relevance at macrocrack stages, suggesting the need for adaptive monitoring strategies. Cross-validation confirmed lower loss variance in the Transformer model, indicating superior generalization. This combined analysis highlights the trade-offs between physics consistency and modeling flexibility, and supports hybrid ML integration for interpretable, robust SHM. • ML models compared Physics-Informed Neural Networks and Transformers for SHM of self-sensing cementitious composite. • The transformer model achieved high accuracy, while PINN ensured physics-consistent predictions with a narrower error margin. • Voltage drop increased during crack initiation and tended to decrease as macro cracks formed. • Strain loses relevance in later crack stages, indicating the need for stage-specific monitoring strategies. • Environmental parameters (temperature, humidity) significantly influence sensing response.