TY - JOUR
T1 - Convolutional neural network-based damage detection method for building structures
AU - Oh, Byung Kwan
AU - Glisic, Branko
AU - Park, Hyo Seon
N1 - Funding Information:
This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korea government (Ministry of Science, ICT & Future Planning, MSIP) (NRF-2021R1A2C33008989 and No. 2018R1A5A1025137).
Funding Information:
This work was supported by a National Research Foundation of Korea (NRF) grant funded by the Korea government (Ministry of Science, ICT & Future Planning, MSIP) (NRF-2021R1A2C33008989 and No.
Publisher Copyright:
Copyright © 2021 Techno-Press, Ltd.
PY - 2021/6
Y1 - 2021/6
N2 - This study presents a damage detection method based on modal responses for building structures using convolutional neural networks (CNNs). The modal responses used in the method are obtained from the dynamic responses, which are measured in a building structure under ambient excitations; these are then transformed to a modal participation ratio (MPR) value for a measuring point and mode. As modal responses vary after damages in the structures, the MPR for a specific location and mode also changes. Thus, in this study, MPR variations, which can be obtained by comparing the MPRs of damaged and healthy structures, are utilized for damage detection without the need for identification of modal parameters. Since MPRs are derived for the number of measuring points (N) in the structure as well as the same number of modes (N), the MPRs and MPR variations can be arranged as an N × N matrix. This low-dimensional MPR variations set is used as the input map of the presented CNN architecture and information about damage locations and severities of the target structure is set as the output of the CNN. The presented CNN is trained for establishing the relationship between MPR variations and damage information and utilized to estimate the damage. The presented damage detection method is applied to numerical examples for two multiple degrees of freedoms and a three-dimensional ASCE benchmark numerical model. Training datasets created from damage scenarios assuming changes in the stiffness are used to train the CNN and the performance of this CNN is verified. Finally, this study examines how variations in the operator size and number of layers in the CNN architecture affect the damage detection performance of CNNs.
AB - This study presents a damage detection method based on modal responses for building structures using convolutional neural networks (CNNs). The modal responses used in the method are obtained from the dynamic responses, which are measured in a building structure under ambient excitations; these are then transformed to a modal participation ratio (MPR) value for a measuring point and mode. As modal responses vary after damages in the structures, the MPR for a specific location and mode also changes. Thus, in this study, MPR variations, which can be obtained by comparing the MPRs of damaged and healthy structures, are utilized for damage detection without the need for identification of modal parameters. Since MPRs are derived for the number of measuring points (N) in the structure as well as the same number of modes (N), the MPRs and MPR variations can be arranged as an N × N matrix. This low-dimensional MPR variations set is used as the input map of the presented CNN architecture and information about damage locations and severities of the target structure is set as the output of the CNN. The presented CNN is trained for establishing the relationship between MPR variations and damage information and utilized to estimate the damage. The presented damage detection method is applied to numerical examples for two multiple degrees of freedoms and a three-dimensional ASCE benchmark numerical model. Training datasets created from damage scenarios assuming changes in the stiffness are used to train the CNN and the performance of this CNN is verified. Finally, this study examines how variations in the operator size and number of layers in the CNN architecture affect the damage detection performance of CNNs.
KW - Convolutional neural network
KW - Damage detection
KW - Dynamic response
KW - Modal participation ratio
KW - Structural health monitoring
UR - http://www.scopus.com/inward/record.url?scp=85108262147&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85108262147&partnerID=8YFLogxK
U2 - 10.12989/sss.2021.27.6.903
DO - 10.12989/sss.2021.27.6.903
M3 - Article
AN - SCOPUS:85108262147
SN - 1738-1584
VL - 27
SP - 903
EP - 916
JO - Smart Structures and Systems
JF - Smart Structures and Systems
IS - 6
ER -