- Research Article
1
- 10.1088/1361-665x/ae515f
An augmented CA-2DCNN framework: achieving high-precision damage identification in steel frame structures
- Mar 01, 2026
- Smart Materials and Structures
- Zhe-Feng Chen + 4 more +4
To enhance the accuracy of damage identification for steel frame structures in scenarios with insufficient data, a new damage identification framework termed augmented CA- two-dimensional convolutional neural network (2DCNN) is proposed. It integrates multi-synchrosqueezing transform (MSST), improved deep convolutional generative adversarial network (IDCGAN), and 2DCNN embedded with coordinate attention (CA) mechanism. Firstly, MSST is employed to transform the measured acceleration response signals into two-dimensional time–frequency images, thereby constructing an original precise feature dataset. Secondly, to address the potential issue of insufficient sample volume, the IDCGAN model is constructed to expand the original feature dataset. Finally, the 2DCNN model embedded with a CA mechanism is established to dynamically extract damage features from the time–frequency images. To validate the effectiveness and accuracy of the proposed augmented CA-2DCNN model, a numerical case of the IASC-ASCE structural health monitoring benchmark model and two experimental tests on a Qatar University Grandstand Simulator and a scaled four-story steel frame structure are investigated. The results demonstrate that the augmented CA-2DCNN model outperforms both the MSST-2DCNN and MSST-CA-2DCNN models in terms of damage identification accuracy across various steel frame structure datasets with various damage locations and severities. Notably, the applied IDCGAN model effectively addresses the challenge of inadequate data samples, whereas the embedded CA mechanism significantly improves the traditional 2DCNN’s capacity to focus on and discriminate critical damage features within time–frequency images. Furthermore, when subjected to identical experimental conditions, in comparison with the current mainstream convolutional network model of ConvNeXt, the augmented CA-2DCNN model achieves an average increase of approximately 15%–20% in damage identification accuracy. Even when confronted with a relatively small sample size, it remains capable of maintaining stable performance. Consequently, the application potential of the proposed method in steel frame structure health monitoring tasks is confirmed.
Read more