Convolutional autoencoder-based framework for damage localization under variable temperature

Confounding factors such as variable temperature have an impact on Lamb wave behaviour, affecting the accuracy of damage localization methods based on such waves. In this study, an innovative approach to Lamb wave prediction based on convolutional autoencoders (CAEs) is proposed and applied to an experimental dataset consisting of Lamb wave acquisitions on a Carbon Fiber Reinforced Polymer (CFRP) plate under varying temperatures. Leveraging an experimental dataset of Lamb wave signals acquired from a CFRP plate at different temperatures, this research focuses on utilizing CAEs to enhance the accuracy and reliability of damage localization. This algorithm extracts critical features from Lamb wave data, effectively recognizing subtle wave properties variations, thus significantly improving the precision of damage localization. Two different architectures of CAEs were evaluated. One which uses the temperature value as a direct input into the latent space of the autoencoder, and another that does not process the temperature value. This analysis was performed to demonstrate the actual impact of the temperature information on prediction accuracy and, furthermore, the accuracy of the CAEs at predicting Lamb wave signals for temperatures outside of their training dataset. The results obtained demonstrate that the inclusion of the temperature information into the autoencoder architecture not only increased its accuracy for temperatures within its training dataset but also increased its robustness with regards to temperature variations, displaying better performance at predicting Lamb wave signals for temperatures outside of its training dataset. The algorithm proposed here presents a way forward for increasing the robustness and reliability of damage localization methods based on Lamb waves.