An Unsupervised Graph Learning Network for Early Damage Intelligent Detection of CFRP

Intelligent early damage recognition technology plays a crucial role in enhancing the reliability and robustness of structural health monitoring (SHM) for carbon fiber-reinforced plastics (CFRP). Early damage recognition techniques are classified into supervised and unsupervised paradigms. The performance of supervised learning-based intelligent early damage recognition models depends heavily on the availability of high-quality labeled data. However, obtaining such data is challenging due to the time-consuming, costly, and impractical nature of acquiring comprehensive damage information for diverse environments and operating conditions. Therefore, developing unsupervised early damage recognition models, which do not rely on labeled data, is essential. This article proposes an unsupervised graph learning network (UGLN) model. The model comprises three main steps. First, the unsupervised dual-order similarity GraphSAGE (DSGraphSAGE) module extracts both local and global structural information between samples to obtain distinguishable damage features. Second, these extracted features are input into an improved sparrow search algorithm-based affinity propagation (SSA-AP) clustering algorithm to group the damage states of different samples. Third, the dissimilarity between each sample’s features and the clustering center representing normal state samples is calculated to generate an evaluation curve. This curve, combined with the clustering results, facilitates accurate identification of the early damage state of CFRP. The proposed method’s effectiveness is demonstrated using a publicly available CFRP accelerated life experimental dataset, capable of identifying early damage samples with 100% accuracy.