Semi-supervised generative adversarial network (SGAN) for damage detection in a composite plate using guided wave responses

Abstract

Data-driven structural health monitoring is becoming popular for its easy application to complex structures. Machine learning or deep learning models are crucial for predicting damage in structures, but their performance depends on the size of the training dataset. Obtaining data can be time-consuming, expensive, and difficult. This study is on the development of a semi-supervised generative adversarial network (SGAN) model that can be trained on fewer samples. The SGAN model is trained and tested on a benchmark dataset from the openguidedwaves (OGW) platform (Moll et al., 2019). The OGW dataset is a collection of Lamb wave interactions with a fiber reinforced composite plate under pristine and damaged conditions. In this study, three SGAN models have been developed using different numbers of supervised training samples and their performance has been evaluated. The developed SGAN models are also compared to the most commonly used deep learning (DL) algorithm, convolutional neural network (CNN) based models, and ResNet autoencoder-based transfer learning models to provide a more comprehensive performance assessment. Compared to CNN models and ResNet autoencoder-based transfer learning (TL) models, SGAN models demonstrated remarkable generalization abilities in detecting damage within composite plates. The performance of all models (SGAN, CNN, and TL) improved as the number of supervised samples increased when compared to their previous versions with fewer supervised samples. When accuracy is used as the performance metric, the top-performing SGAN model outperformed the leading CNN model by 24.62% and the best transfer learning (TL) model by 10.96%.