Study on structural damage identification based on two-dimensional images and improved convolutional neural networks

Abstract

To reduce the cost of structural damage detection, reduce the number of measurement points and samples, and address the issues of low recognition accuracy, weak noise robustness, and poor reliability in traditional deep learning models in the field of structural damage identification. This paper proposes a structural damage identification method that combines a convolutional neural network incorporating the SE attention mechanism and a frequency domain grayscale image (FDGSI). Firstly, a finite element model of a slender beam is established, and vibration signals containing damaged information are collected. Numerical encoding is used to convert the vibration signals into grayscale images (GSIs) using a two-dimensional discrete Fourier transform (2D-DFT), which transform the GSIs from the spatial domain to the frequency domain. Secondly, the SE-CNN and test models are built, and structural damage identification studies are conducted on the datasets by combining different models. The effects of noise data and cross-point data are also analyzed. Finally, the proposed model was validated using the long steel beam dataset and the grandstand simulator dataset from prior studies. The study results show that the model achieved over 99.5% recognition accuracy in the original data of three datasets. In identifying noise data, it achieved recognition accuracies of 99.10%, 98.9%, and over 99%, respectively. In identifying cross-point data, recognition accuracy is achieved at 97.60%, 96.26%, and 99.09%, respectively. These results demonstrate that the proposed method has specific generalization, practicality, and noise-resistant robustness.