A Time-Frequency-Based Data-Driven Approach for Structural Damage Identification and Its Application to a...

Area of Science:

• Structural engineering
• Machine learning applications
• Signal processing

Background:

• Traditional structural damage identification relies on manual feature extraction, often leading to suboptimal performance.
• Deep learning methods, like Convolutional Neural Networks (CNNs), offer automated feature extraction for improved accuracy in Structural Health Monitoring (SHM).

Purpose of the Study:

• To develop an advanced data-driven approach for structural damage identification in complex structures.
• To enhance the effectiveness of existing methods using time-frequency analysis and deep learning.

Main Methods:

• Converted structural acceleration signals into 2D images using the Gram Angle Difference Field (GADF).
• Employed Convolutional Neural Networks (CNNs), specifically ResNet, to extract features and classify damage from image data.
• Conducted experimental validation on a cable-stayed bridge model under moving vehicle loads.

Main Results:

• The ResNet model demonstrated superior performance in damage identification accuracy and convergence speed compared to four other traditional networks.
• The proposed method accurately identified damage on bridges using limited sensors on the deck.
• Prediction accuracy decreased from 86.63% to 62.5% as Signal-to-Noise Ratio (SNR) dropped from 20 dB to 2.5 dB, indicating robustness.

Conclusions:

• The time-frequency-based, data-driven approach using CNNs (ResNet) is effective for structural damage identification.
• The method shows significant potential for real-world applications on bridges, even with environmental noise and limited sensor data.