This study addresses the challenge of extracting and classifying damage features in steel structures under multiple-damage scenarios. It aims to develop an enhanced deep learning framework for processing vibration signals, thereby advancing structural health monitoring (SHM) and enabling more accurate and intelligent damage identification.
A novel one-dimensional dense connection convolutional network model enhanced with a convolutional block attention module (1D-CDNet) is proposed. The method adapts DenseNet by replacing its two-dimensional convolutional kernels with one-dimensional kernels. This modification constructs a one-dimensional densely connected convolutional network (1D-DenseNet) specifically designed for vibration signal analysis. Furthermore, an improved convolutional block attention module (ICBAM) is integrated to adaptively emphasize informative channels and temporal instances while maintaining overall network stability. The proposed 1D-CDNet architecture is constructed by sequentially stacking the 1D convolutional layer, 1D-DenseNet and ICBAM. Experiments were conducted using I-beam vibration signals and the IASC-ASCE benchmark dataset.
The results demonstrate that 1D-CDNet achieves high accuracy in identifying damage in I-beam frame structures, with accuracies of 99.57% and 99.21% under single-damage and multiple-damage scenarios, respectively. Experiments conducted on the IASC-ASCE benchmark dataset confirm its effectiveness and strong generalization capability. Compared with the baseline 1D-DenseNet, 1D-CDNet improves classification accuracy by 1.41%. Overall, 1D-CDNet outperforms the comparative models, while the proposed ICBAM module reduces information loss, enhances robustness and improves performance under multiple-damage conditions.
This study introduces a novel deep learning architecture for identifying damage in steel structures using vibration signals. The proposed model adapts DenseNet for one-dimensional vibration signal analysis and integrates an enhanced attention mechanism, thereby improving the accuracy of steel structure damage identification, particularly under complex multiple-damage scenarios.
