The rapid development of structural health monitoring (SHM) technology has greatly accelerated the data-driven methods for structural damage identification. However, the scarcity of accurately labeled data remains a primary limitation for supervised models in predicting structural conditions. This study develops an unsupervised learning framework integrating ConvNeXt model with the Deep Subdomain Adaptation Network (DSAN) to conduct feature extraction and cross-domain alignment. The proposed framework achieves precise structural damage identification with unlabeled datasets by utilizing the powerful feature extraction capability of the ConvNeXt model and the efficient cross-domain alignment of DSAN. The monitored structural vibration signals were initially converted into time–frequency representations via Gramian Angular Difference Field (GADF) to construct image datasets. Subsequently, the ConvNeXt-DSAN was utilized for feature extraction, feature alignment, and feature classification. The proposed framework was validated by an experimental study on a truss model. The numerical result demonstrates the effectiveness of the proposed method, which achieves an average target-domain recognition accuracy of 84.56%. In contrast, the conventional Convolutional Neural Networks (CNN)-based DSAN models, including AlexNet, GoogLeNet, VGG16, and ResNet, are 75.43 %, 79.12 %, 77.57 %, and 80.78 %, respectively. Furthermore, the proposed approach exhibits faster convergence and greater computational efficiency. The ConvNeXt-DSAN framework addresses key challenges in SHM tasks for effective damage identification based on unlabeled data. It provides a promising foundation for extending unsupervised learning approaches to intelligent operation and maintenance of large-scale civil infrastructures.