An adversarial transformer for anomalous lamb wave pattern detection.

Area of Science:

• Structural Health Monitoring
• Wave Propagation Analysis
• Machine Learning for Engineering

Background:

• Lamb waves are crucial for structural health monitoring, necessitating advanced data analysis techniques.
• Existing methods for Lamb wave analysis face challenges in reliably detecting anomalous patterns.

Purpose of the Study:

• To develop an unsupervised Adversarial Transformer model for anomalous Lamb wave pattern detection.
• To enhance defect detection capabilities in structural health monitoring applications.

Main Methods:

• Utilized a hybrid PZT-scanning laser Doppler vibrometer (SLDV) to generate spatiotemporal Lamb wave images.
• Implemented an unsupervised Adversarial Transformer model with global and local attention mechanisms trained adversarially.
• Introduced a segment replacement strategy to improve the extraction of normal wave textural features.

Main Results:

• The Adversarial Transformer model achieved 97.1% accuracy in anomalous wave pattern detection.
• Global attention effectively reconstructed normal data while highlighting anomalies.
• Adversarial training with local attention boosted the performance of global attention.

Conclusions:

• The proposed Adversarial Transformer model offers superior performance for anomalous Lamb wave pattern detection compared to benchmark models.
• The combination of global and local attention within an adversarial framework is key to the model's effectiveness.
• The segment replacement strategy further enhances the model's ability to distinguish normal from anomalous wave patterns.