A Novel Unsupervised Structural Damage Detection Method Based on TCN-GAT Autoencoder.

Area of Science:

• Structural Engineering
• Artificial Intelligence
• Signal Processing

Background:

• Structural damage detection is vital for long-term safety and durability of infrastructure.
• Existing methods often fail to capture complex spatiotemporal correlations in multi-sensor data.
• This limitation hinders the full utilization of structural dynamic evolution and spatial relationships.

Purpose of the Study:

• To develop an unsupervised damage detection method that effectively integrates spatiotemporal features from structural vibration data.
• To address the limitations of existing methods by explicitly modeling both temporal dependencies and spatial correlations.
• To create a robust framework suitable for real-time structural health monitoring.

Main Methods:

• Proposes a novel autoencoder model, TCNGAT-AE, combining Temporal Convolutional Networks (TCN) for temporal feature extraction and Graph Attention Networks (GAT) for spatial relationship modeling.
• Employs an "offline training-online detection" strategy using only healthy state data for training.
• Utilizes reconstruction error from the autoencoder as the indicator for structural damage.

Main Results:

• The TCNGAT-AE model demonstrated effective damage detection across different structures (concrete bridge, steel bridge) and excitation types (ambient, vehicle load).
• Explicit spatiotemporal feature modeling significantly improved detection performance and anomaly detection rates compared to models using only temporal or spatial analysis.
• Ablation studies confirmed the effectiveness of the spatiotemporal fusion mechanism.

Conclusions:

• The TCNGAT-AE method provides a powerful tool for unsupervised structural damage detection, capable of handling complex engineering environments.
• The end-to-end framework processes raw vibration signals directly, enabling practical and near-real-time monitoring applications.
• This approach can be integrated into real-time monitoring systems for critical structures, enhancing overall structural safety and maintenance strategies.