科学研究
学术论文

A modified multiscale semantic segmentation network accounting for multi-level seismic damage features of PC structure

来源:   作者:  发布时间:2024年05月01日  点击量:

A modified multiscale semantic segmentation network accounting for multi-level seismic damage features of PC structure

Dianyou Yu , Zheng He , Ling Ma

Abstract

Compared with cast-in-place reinforced concrete structures, precast concrete (PC) structures have a higher possibility of concentrated plasticity development in connections while experiencing severe damage caused by strong earthquakes. How to rationally select the prominent damage features is crucial for conducting damage assessment. This article modifies an advance semantic segmentation method, TransUNet, which automatically addresses the extraction of multiscale damage features from images of PC structure. Based on transfer learning, training, validation, and testing processes are conducted under a self-built dataset. The encoder-decoder network is improved by introducing the Transformer blocks to enhance the self-attention for capturing multiple contextual prior knowledge or global knowledge, the pre-trained ResNet50 blocks are employed to address the scarcity of seismic dataset and generate generic features, as well as the multiscale fusing blocks and loss function for effectively balancing attention across the multiscale damage features of PC structure. Moreover, the effectiveness of the components is demonstrated through an ablation study. The test results demonstrate that the modified network achieves acceptable evaluation metrics and consistent performance. The segmentation accuracy of small-scale damage features exhibits an acceptable improvement and adequately fulfills the demand for damage assessment of PC structure. Additionally, the network demonstrates excellent segmentation ability on multi-level damage features and exhibits robust performance against various noises.

Keywords:
Precast concrete structure;
Multiscale semantic segmentation;
Transformer;
Multiscale loss function;
Deep learning

https://www.sciencedirect.com/science/article/pii/S2352710223017801