Ultra-high-performance concrete-filled steel tubular (UHP-CFST) members exhibit excellent overall performance; however, the opaque outer steel tube completely prevents direct observation of the damage evolution within the UHPC core. With the continuing development of deep and distant offshore regions, engineering structures are increasingly exposed to extreme loading conditions, which can cause irreversible damage to the core concrete of CFST members. Therefore, the introduction of structural health monitoring (SHM) techniques for accurate identification of structural damage severity is particularly important for these members. Acoustic emission (AE) technology, owing to its in situ passive monitoring capability and sensitivity to microscopic defects, is well suited to such harsh conditions and has become an ideal nondestructive evaluation tool for this type of member. However, although AE enables dynamic monitoring, conventional parameter-based analysis methods, such as the Ib-value and hit count, still rely heavily on manual experience and have limited capability in representing the complex nonlinear relationships embedded in AE signals, which makes it difficult to fully extract the latent damage information contained in high-dimensional AE data. In recent years, deep learning (DL) has been widely applied in pattern recognition, image classification, and regression tasks. Compared with conventional methods, DL methods can automatically extract high-dimensional features and establish nonlinear mappings among complex data, and have therefore attracted increasing attention in the field of structural health monitoring. Against this background, axial compression tests were conducted on 27 UHP-CFST short columns with synchronous AE data acquisition, and an intelligent classification framework was developed to identify damage stages in the specimens.

A total of 27 UHP-CFST short columns with an outer diameter of 152 mm and a height of 456 mm were tested under concentric axial compression. The test matrix included three steel grades (Q235, Q355, and Q420) and three wall thicknesses (4, 6 mm, and 8 mm), and all specimens were filled with UHPC and three replicate specimens were prepared for each group. The tests were conducted under displacement control, with a loading rate of 1.0 mm/min in the initial stage and 2.0 mm/min after the peak load was reached. The applied load and axial displacement were recorded synchronously by the built-in data acquisition module of the testing machine and external high-precision linear variable differential transformers (LVDTs), respectively. Meanwhile, a 16-channel MISTRAS acoustic emission monitoring system developed by Physical Acoustics Corporation (PAC), USA, was used to monitor crack development in the UHPC core throughout the test.

Based on the evolution of normalized cumulative hits (NCH) as a function of load percentage, the damage process was divided into three stages. AE feature parameters were subsequently extracted, and minority-class data were augmented using a Wasserstein generative adversarial network with gradient penalty (WGAN-GP) to mitigate class imbalance. Based on the augmented data, a damage-stage identification model based on a one-dimensional residual network (1D-ResNet) was further established to classify different damage stages intelligently.

The AE samples generated by WGAN-GP exhibited high similarity to the original samples in terms of both statistical characteristics and feature-space distribution. For Stage I damage identification, the developed ResNet model achieved a precision of 88.4%, a recall of 95.0%, and an F1-score of 91.6%. These results indicated that the model was able to identify most Stage I samples, with relatively few missed detections. The high recall suggested that the model was highly sensitive to the initial damage stage and could effectively capture early damage features. However, the relatively lower precision indicated that some Stage II samples were misclassified as Stage I, implying the presence of a limited number of false positives. For Stage II damage identification, the model achieved a precision of 92.7%, a recall of 95.0%, and an F1-score of 93.8%. All three metrics remained at a high level, and the precision and recall were relatively well balanced. These results indicated that the model was able to effectively extract the AE characteristics of Stage II and accurately distinguish them from those of the other stages, thereby yielding the best overall identification performance. For Stage III damage identification, the model achieved a precision of 100%, a recall of 90.0%, and an F1-score of 94.7%. These results indicated that the model exhibited extremely high identification accuracy for Stage III. When a sample was classified as belonging to Stage III, no misclassification occurred, indicating the absence of false positives. However, some true Stage III samples were assigned to other stages, which led to a slightly lower recall. The confusion matrix further showed that the model exhibited strong identification capability for all three damage stages. Among them, the classification performance for Stages II and III was particularly good. A small number of misclassifications were observed between Stages I and II, which could be attributed mainly to the continuity of the damage evolution process, as some late Stage I samples and early Stage II samples exhibited a certain degree of overlap in the AE feature space.

In addition, the comparison of identification results among different models showed that the ResNet model achieved high classification accuracy across all three damage stages, with overall performance markedly superior to that of the SVM, MLP, and CNN models. Overall, the results demonstrated that the ResNet model constructed on the basis of the augmented dataset could effectively learn the damage evolution patterns embedded in the AE features and achieve accurate identification of different damage stages in UHP-CFST short columns under axial compression.

The data generated by WGAN-GP show high consistency with the original data in terms of both statistical characteristics and the distribution in high-dimensional feature space, indicating that the overall distribution trend and the main feature structure of the AE data are well preserved. On this basis, a three-stage damage identification model based on ResNet is established and further compared with support vector machine (SVM), multilayer perceptron (MLP), and convolutional neural network (CNN) models. The results show that the F1-scores of the ResNet model for Stages I, II, and III reach 91.6%, 93.8%, and 94.7%, respectively, demonstrating overall identification performance superior to that of the SVM, MLP, and CNN models. Therefore, the proposed method combining WGAN-GP and ResNet effectively improves the intelligent identification of axial compression damage stages in UHP-CFST short columns and provides a new perspective for AE-based structural damage identification.