LambNet-T: A lightweight path-conditional transformer autoencoder for temperature-aware baseline learning in...

Area of Science:

• Structural Health Monitoring (SHM)
• Acoustic Wave Propagation
• Machine Learning for Engineering

Background:

• Reliable Structural Health Monitoring (SHM) using Lamb waves is challenged by temperature variations impacting baseline data.
• Accurate baseline selection is crucial for effective damage detection and localization in SHM systems.
• Existing methods often struggle with temperature compensation and multi-path data integration.

Purpose of the Study:

• To develop a temperature-aware baseline learning method for multi-path Lamb-wave SHM.
• To introduce LambNet-T, a lightweight Transformer-based autoencoder for efficient and robust baseline selection.
• To improve the diagnostic accuracy and temperature robustness of SHM systems.

Main Methods:

• Developed LambNet-T, a path-conditional Transformer-based autoencoder utilizing Attention Pooling (AP) for contextual embeddings.
• Implemented a Cosine Similarity (CS) with a Median-based evaluation strategy for robust baseline selection.
• Employed quadratic interpolation for data augmentation and used limited baseline temperatures (-10 to +50 °C) reflecting practical constraints.

Main Results:

• LambNet-T exhibited superior training efficiency compared to a convolutional autoencoder (CAE-GAP).
• The Median of highest path-specific CS values effectively identified the optimal temperature-compensated baseline.
• Achieved high precision (R² = 0.994 ± 0.001), outperforming CAE-GAP and conventional Optimal Baseline Selection (OBS).
• Integration reduced impact location errors to 4.12 mm and incorporated a statistical filter for uncertainty management.

Conclusions:

• LambNet-T provides a highly accurate and temperature-robust solution for multi-path Lamb-wave SHM baseline selection.
• The proposed method significantly enhances diagnostic performance and efficiency in practical SHM applications.
• Openly available datasets ensure reproducibility and facilitate further research in temperature-aware SHM.