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

Reliable Lamb-wave-based Structural Health Monitoring (SHM) depends on accurate baseline selection under varying temperatures. This study presents LambNet-T, a lightweight path-conditional Transformer-based autoencoder for temperature-aware baseline learning across multiple transducer paths. LambNet-T employs Attention Pooling (AP) to generate contextual embeddings and enables robust baseline selection using Cosine Similarity (CS) with a Median-based evaluation strategy, improving diagnostic accuracy and temperature robustness in multi-path Lamb-wave SHM. Experiments on a composite panel over -10 to +50 °C used only four baseline temperatures to reflect practical constraints, with quadratic interpolation for data augmentation. LambNet-T demonstrated significantly higher training efficiency than a convolutional autoencoder (CAE-GAP). During inference, the Median of the highest path-specific CS values identified the optimal temperature-compensated baseline. The method achieved high precision (R² = 0.994 ± 0.001), outperforming both CAE-GAP and conventional Optimal Baseline Selection (OBS). Integration with an existing damage localization framework reduced impact location errors to as low as 4.12 mm. A conservative statistical filter, based on baseline selection variability, was applied to manage uncertainty. All experimental datasets are openly available for reproducibility.