Lamb wave-based damage localization in composite laminates using LSTM networks trained with improved loss functions

Abstract Composite materials have been extensively employed in industrial fields due to their superior properties. Structural health monitoring and non-destructive testing techniques based on Lamb waves have been used for damage detection in composite laminates. Recently, deep learning algorithms have attracted attention and been utilized to extract damage features from Lamb waves. To fully exploit damage characteristics in Lamb waves and improve the performance of trained models, an long short-term memory (LSTM)-based method for damage localization and novel loss functions for model training are proposed in this study. Firstly, an LSTM-based network is established to extract features from sequential Lamb wave signals. Then, numerical simulated signals are generated to construct a training dataset. During model training, location-based and sparsity-based loss functions are designed for the optimization of network parameters. Since the proposed loss functions have physical interpretability to some extent, they are expected to accelerate the convergence of model training and increase the generalization ability of trained networks. The experiment is implemented on a carbon fiber reinforced plastics plate with simulated damage. The results demonstrate that the networks trained with numerical data can localize damage in the experimental specimen with good generalization ability, which validates the effectiveness of the proposed approach for damage localization in composite laminates.