A Novelty Detection Approach for Tendons of Prestressed Concrete Bridges Based on a Convolutional Autoencoder and...

Area of Science:

• Structural Engineering
• Bridge Health Monitoring
• Non-Destructive Testing

Background:

• Prestressed concrete (PSC) bridge safety relies heavily on the integrity of prestressed tendons.
• Real-time detection of tendon damage in PSC bridges remains a significant challenge.

Purpose of the Study:

• To propose and evaluate a novel approach for detecting prestressed tendon damage in PSC bridges using a convolutional autoencoder (CAE).
• To assess the effectiveness of the CAE method under various damage severities and measurement error levels.

Main Methods:

• Utilized simulation data from nine accelerometers to train and test the CAE model.
• Investigated the CAE's performance for both multi-vehicle and single-vehicle scenarios.
• Analyzed detection accuracy across different damage severities (100%, 75%, 50%) and error levels (0%, 5%, 10%).

Main Results:

• CAE achieved 79.5%–85.8% accuracy for multi-vehicle scenarios with severe damage (100%, 75%) and no error.
• Accuracy dropped to 69.4%–73.3% for moderate (50%) multi-vehicle damage with 5-10% error.
• CAE demonstrated high accuracy (90.1%–95.1%) for single-vehicle scenarios across all damage severities and error levels.

Conclusions:

• The CAE approach is effective for detecting severe tendon damage in PSC bridges under multi-vehicle conditions.
• For moderate damage, the CAE's effectiveness in multi-vehicle scenarios is limited, especially with measurement errors.
• Single-vehicle acceleration data enables a highly accurate and robust CAE-based method for PSC bridge tendon damage detection, even with significant errors.