Damage detection of steel frame structure based on IECA-VAE and IECANet under imbalanced data

The structural integrity of steel frame systems is paramount to the safety and stability of large-scale civil infrastructure. Consequently, neural network models have emerged as a widely adopted methodology for monitoring the evolution of structural damage. In the case of imbalanced data, the efficiency of such neural network-based methods is greatly reduced, leading to a decrease in the accuracy of damage identification. To address the challenge of damage detection of steel frame structures under imbalanced data conditions, this article integrates the Inception module and Efficient Channel Attention (ECA) mechanism to develop the IECA module. Furthermore, it proposes the IECA-VAE network for sample generation and the IECANet framework for damage detection. Tests conducted on the experimental datasets of the American Society of Civil Engineers (IASC-ASCE) benchmark model and Qatar University grandstand simulator (QUGS) demonstrate the following results. For the data generated by IECA-VAE, the root mean square error between the generated data and real signals is 0.0018 and 0.0019, respectively; the spectral correlation coefficients reach 0.9933 and 0.9913, respectively. Additionally, both the spectral mean square error and the log spectral distance of IECA-VAE are the lowest among the three comparative methods, indicating its highest generation quality and spectral distortion. Under normal data conditions, IECANet achieves over 94% (IASC-ASCE) and 95% (QUGS) in multiple metrics; in the R0G2 dataset augmented by IECA-VAE, the mean values of some metrics even reach 100% and 99.94%. Under small-sample conditions, its average recognition accuracy still exceeds 91% and 93%, respectively, and further reaches 95% and 99.5% on the R0G2 dataset. This indicates that the method proposed in this article can effectively handle the precise damage detection problem of steel frame structures under data imbalance conditions. Meanwhile, the proposed methods also have the potential to be extended to damage detection in other structural forms.