TY - JOUR
T1 - Knowledge transfer for structural damage detection through re-weighted adversarial domain adaptation
AU - Wang, Xiaoyou
AU - Xia, Yong
N1 - Funding Information:
This study was supported by the Key-Area Research and Development Program of Guangdong Province (Project No. 2019B111106001), RGC-GRF (Project No. 15201920), and PolyU Project of Strategic Importance (Project No. 1-ZE1F). The first author X.Y. Wang is grateful for the valuable discussion with Dr. Jinyang Jiao from the Beihang University, China.
Funding Information:
This study was supported by the Key-Area Research and Development Program of Guangdong Province (Project No. 2019B111106001 ), RGC-GRF (Project No. 15201920 ), and PolyU Project of Strategic Importance (Project No. 1-ZE1F ). The first author X.Y. Wang is grateful for the valuable discussion with Dr. Jinyang Jiao from the Beihang University, China.
Publisher Copyright:
© 2022 Elsevier Ltd
PY - 2022/6/1
Y1 - 2022/6/1
N2 - Deep learning (DL) techniques have been developed for structural damage detection by training the network to dig damage-sensitive features from big data. However, most techniques only perform well on datasets with the same distribution as the training data. The network needs to be re-trained by re-collecting labeled data when the environmental conditions or structural sizes change. This limits the application of DL techniques to damage detection of practical structures, since many bridges may have the same topology but different sizes, whereas re-collecting labeled damaged data is expensive and often infeasible in structural health monitoring. A re-weighted adversarial domain adaptation (RADA) method is developed to generalize the network trained on one structure to others without re-collecting the labeled data. As damage is irreversible, the damage cases in structures may be different. Considering the inconsistent label spaces between the source and target domains, a weight parameter is introduced to improve the importance of the shared label space in the DA process. The RADA network learns damage-sensitive and domain-invariant features for the damage detection of the new structure by training the generator and two classifiers in an adversarial manner. The proposed method is applied to two types of knowledge transfer, namely, from one structure to the other with different sizes and from a numerical model to an experimental structure. Examples show that the RADA network significantly improves the classification accuracy in transfer learning problems with inconsistent label spaces, as compared with the networks without DA or without the re-weighting mechanism. The method can also be extended to other unsupervised classification problems with label scarcity.
AB - Deep learning (DL) techniques have been developed for structural damage detection by training the network to dig damage-sensitive features from big data. However, most techniques only perform well on datasets with the same distribution as the training data. The network needs to be re-trained by re-collecting labeled data when the environmental conditions or structural sizes change. This limits the application of DL techniques to damage detection of practical structures, since many bridges may have the same topology but different sizes, whereas re-collecting labeled damaged data is expensive and often infeasible in structural health monitoring. A re-weighted adversarial domain adaptation (RADA) method is developed to generalize the network trained on one structure to others without re-collecting the labeled data. As damage is irreversible, the damage cases in structures may be different. Considering the inconsistent label spaces between the source and target domains, a weight parameter is introduced to improve the importance of the shared label space in the DA process. The RADA network learns damage-sensitive and domain-invariant features for the damage detection of the new structure by training the generator and two classifiers in an adversarial manner. The proposed method is applied to two types of knowledge transfer, namely, from one structure to the other with different sizes and from a numerical model to an experimental structure. Examples show that the RADA network significantly improves the classification accuracy in transfer learning problems with inconsistent label spaces, as compared with the networks without DA or without the re-weighting mechanism. The method can also be extended to other unsupervised classification problems with label scarcity.
KW - Adversarial domain adaptation
KW - Deep learning
KW - Inconsistent label space
KW - Knowledge transfer
KW - Structural damage detection
UR - http://www.scopus.com/inward/record.url?scp=85125946706&partnerID=8YFLogxK
U2 - 10.1016/j.ymssp.2022.108991
DO - 10.1016/j.ymssp.2022.108991
M3 - Journal article
AN - SCOPUS:85125946706
SN - 0888-3270
VL - 172
JO - Mechanical Systems and Signal Processing
JF - Mechanical Systems and Signal Processing
M1 - 108991
ER -