Bridge Damage Identification Using Deep Neural Networks on Time-Frequency Signals Representation.

Area of Science:

• Civil Engineering
• Structural Health Monitoring
• Signal Processing
• Machine Learning

Background:

• Maintaining civil infrastructure requires continuous monitoring of structural integrity.
• Early detection of damage is vital for ensuring the longevity and safety of bridges.
• Existing methods may lack the precision needed for complex, multiclass damage identification.

Purpose of the Study:

• To develop a unique approach for multiclass damage detection in civil structures.
• To classify time-series acceleration responses from bridge accelerometers under various damage scenarios.
• To validate the proposed method using benchmark data from a real-world bridge.

Main Methods:

• Utilizing synchrosqueezing transform (SST) for signal processing of acceleration responses.
• Applying deep learning algorithms, specifically pre-trained 2D convolutional neural networks (CNNs).
• Validating the pipeline on the Z24 bridge benchmark dataset, which includes labeled, real-world damage data.

Main Results:

• The proposed pipeline accurately classifies different types of damage scenarios on a bridge.
• High classification accuracy was achieved by exploiting pre-trained 2D CNNs.
• Simple voting methods were found to further enhance the classification accuracy.

Conclusions:

• The developed method offers a robust and effective solution for multiclass structural damage detection.
• The integration of SST and deep learning provides a powerful tool for bridge health monitoring.
• The approach demonstrates significant potential for practical application in civil infrastructure maintenance.