Cross-Modal Graph Attention for Bridge SHM Data Imputation.

Area of Science:

• Civil Engineering
• Data Science
• Structural Health Monitoring

Background:

• Structural health monitoring (SHM) systems are crucial for bridge maintenance.
• Data loss in SHM systems, due to sensor faults or communication issues, hinders accurate structural assessment.
• Existing single-channel imputation methods fail to capture spatial and cross-modal correlations in heterogeneous data.

Purpose of the Study:

• To develop a collaborative imputation framework for reconstructing missing bridge displacement data.
• To address limitations of conventional methods in handling long-term and continuous data loss.
• To improve the reliability of SHM systems through accurate data imputation.

Main Methods:

• Proposed an integrated framework (ITimeGAN) combining Graph Attention Network (GAT), Modal-Aware Cross-Attention (MACA), and a temporal encoder-decoder architecture.
• Constructed a sensor feature topological graph using Pearson correlation and employed GAT to learn spatial dependencies.
• Utilized MACA for cross-modal information aggregation and a bidirectional/unidirectional LSTM for temporal dependency capture.

Main Results:

• ITimeGAN achieved high accuracy in reconstructing missing displacement data on the Forth Road Bridge dataset.
• Demonstrated effectiveness in both random (10-50%) and continuous long-term missing (1-10 days) scenarios.
• Achieved R² of 0.9950 for longitudinal and 0.9759 for vertical displacement, even with 10 days of missing data.

Conclusions:

• The proposed ITimeGAN framework effectively reconstructs missing bridge displacement data by leveraging spatial and cross-modal correlations.
• The integration of GAT and MACA modules significantly improves imputation accuracy compared to baseline methods.
• The study confirms the effectiveness of a cross-modal collaborative strategy for enhancing SHM system reliability.