Unsupervised learning framework for temperature compensated damage identification and localization in ultrasonic...

Area of Science:

• Structural Health Monitoring
• Ultrasonic Guided Waves
• Machine Learning

Background:

• Current damage localization algorithms for guided wave-based structural health monitoring (GW-SHM) rely on manual features and supervised learning, limiting accuracy with environmental variations and new damage types.
• Deep learning models offer improvements but often require millions of parameters and extensive training data.

Purpose of the Study:

• To propose an unsupervised, temperature-compensated approach for damage identification and localization in GW-SHM systems.
• To reduce the number of trainable parameters by leveraging transfer learning (TL).
• To demonstrate the scalability and adaptability of the framework for diverse applications.

Main Methods:

• Utilizing transfer learning from a convolutional autoencoder (TL-CAE) for unsupervised damage identification and localization.
• Processing raw time-domain signals without pre-processing or material property knowledge.
• Conducting an extensive parametric study to validate the method's feasibility.

Main Results:

• Achieved more accurate damage detection and localization compared to supervised approaches on the Open Guided Waves (OGW) dataset.
• Demonstrated significant robustness to temperature variations.
• Reduced the number of trainable parameters through transfer learning, leveraging time-series similarities across sensor paths.

Conclusions:

• The proposed unsupervised TL-CAE method offers a robust and accurate solution for GW-SHM, overcoming limitations of traditional supervised techniques.
• The framework's ability to use raw data and its parameter reduction make it scalable and adaptable for various materials, structures, and conditions.
• This approach advances the field of structural health monitoring by providing a more efficient and effective damage localization tool.