WaveCLR: Contrastive Learning of Guided Wave Representations for Composite Damage Identification

Damage identification in composite materials is an area of interest within the field of structural health monitoring (SHM). Damage identification in the field not only cares about recognition accuracy but also requires robustness to the variation of the input data. Deep learning methods to analyze guided wave signals collected by ultrasonic testing (UT) for damage identification have been gaining increasing attention in the last decade. However, existing methods for damage identification lack this robustness by failing to consider the input data distribution discrepancy between the training phase (source domain) and the inference phase (target domain). To address this issue, this article proposes a novel domain generalization method named WaveCLR for damage identification in composite materials using ultrasonic guided waves (UGW). WaveCLR is a deep neural network (DNN) model that leverages contrastive learning to account for data discrepancies that typically arise from variations in ultrasonic wave data due to damage characteristics. By maximizing the feature mutual interclass distance and intraclass distance in the source domain, the models' performance in the target domain is improved. The effectiveness of the proposed WaveCLR was validated on two distinct datasets capturing different damage scenarios. The results indicate that our proposed model successfully addresses the domain generalization problem in composite damage identification.