Details
Original language | English |
---|---|
Article number | 117570 |
Journal | Engineering structures |
Volume | 303 |
Early online date | 28 Jan 2024 |
Publication status | Published - 15 Mar 2024 |
Abstract
In this paper, an autoencoder trained with non-standardized time series data and evaluated using covariance-based residuals for generally applicable unsupervised damage localization is investigated. Raw acceleration time series are used as the inputs for the autoencoder to give both these features: no loss of information and exploitation of the full potential of the neural network. When it comes to output-only and unsupervised structural health monitoring (SHM), data-driven models struggle to localize the positions of damage adequately or only work well in a small range of applications. Regarding neural networks, expertise is needed for the neural network dimensioning and understanding of structural dynamics, which increases the difficulty of the task. In order to simplify the process, an automated solution is provided to perform the neural architecture search, and principal component analysis (PCA) is used to find a good choice for the bottleneck dimension. As an extension of the model, the residuals between the original and reconstructed time series are evaluated using the covariance between each input signal and each residual time series, which results in improved indicators for damage localization. We demonstrate the efficiency of the proposed schemes for damage analysis in a series of simulations using a three-mass swinger, in which the autoencoder can localize the damage using varying excitation locations. The covariances’ evaluation indicates that they are more potent than using the reconstruction error. Finally, experimental validation is conducted using vibration test data from a lattice tower called Leibniz University Structure for Monitoring (LUMO) under ambient excitation. For each damage pattern, high sensitivity towards local stiffness is achieved. Additionally, the damage position indicators exhibit a clear decreasing trend as the distance from the damage increases. The autoencoders presented here with non-standardized time series and covariance-based evaluation of residuals lead to increased robustness and sensitivity regarding damage localization.
Keywords
- Autoencoder, Data-driven model, PCA, Structural health monitoring, Unsupervised damage localization
ASJC Scopus subject areas
- Engineering(all)
- Civil and Structural Engineering
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: Engineering structures, Vol. 303, 117570, 15.03.2024.
Research output: Contribution to journal › Article › Research › peer review
}
TY - JOUR
T1 - On using autoencoders with non-standardized time series data for damage localization
AU - Römgens, Niklas
AU - Abbassi, Abderrahim
AU - Jonscher, Clemens
AU - Grießmann, Tanja
AU - Rolfes, Raimund
N1 - Funding Information: The authors gratefully acknowledge the financial support provided by the Federal Ministry for Economic Affairs and Climate Action of the Federal Republic of Germany within the framework of the collaborative research project Grout-WATCH ( FKZ 03SX505B ) and SMARTower ( FKZ 03EE2041C ). All authors approved the version of the manuscript to be published.
PY - 2024/3/15
Y1 - 2024/3/15
N2 - In this paper, an autoencoder trained with non-standardized time series data and evaluated using covariance-based residuals for generally applicable unsupervised damage localization is investigated. Raw acceleration time series are used as the inputs for the autoencoder to give both these features: no loss of information and exploitation of the full potential of the neural network. When it comes to output-only and unsupervised structural health monitoring (SHM), data-driven models struggle to localize the positions of damage adequately or only work well in a small range of applications. Regarding neural networks, expertise is needed for the neural network dimensioning and understanding of structural dynamics, which increases the difficulty of the task. In order to simplify the process, an automated solution is provided to perform the neural architecture search, and principal component analysis (PCA) is used to find a good choice for the bottleneck dimension. As an extension of the model, the residuals between the original and reconstructed time series are evaluated using the covariance between each input signal and each residual time series, which results in improved indicators for damage localization. We demonstrate the efficiency of the proposed schemes for damage analysis in a series of simulations using a three-mass swinger, in which the autoencoder can localize the damage using varying excitation locations. The covariances’ evaluation indicates that they are more potent than using the reconstruction error. Finally, experimental validation is conducted using vibration test data from a lattice tower called Leibniz University Structure for Monitoring (LUMO) under ambient excitation. For each damage pattern, high sensitivity towards local stiffness is achieved. Additionally, the damage position indicators exhibit a clear decreasing trend as the distance from the damage increases. The autoencoders presented here with non-standardized time series and covariance-based evaluation of residuals lead to increased robustness and sensitivity regarding damage localization.
AB - In this paper, an autoencoder trained with non-standardized time series data and evaluated using covariance-based residuals for generally applicable unsupervised damage localization is investigated. Raw acceleration time series are used as the inputs for the autoencoder to give both these features: no loss of information and exploitation of the full potential of the neural network. When it comes to output-only and unsupervised structural health monitoring (SHM), data-driven models struggle to localize the positions of damage adequately or only work well in a small range of applications. Regarding neural networks, expertise is needed for the neural network dimensioning and understanding of structural dynamics, which increases the difficulty of the task. In order to simplify the process, an automated solution is provided to perform the neural architecture search, and principal component analysis (PCA) is used to find a good choice for the bottleneck dimension. As an extension of the model, the residuals between the original and reconstructed time series are evaluated using the covariance between each input signal and each residual time series, which results in improved indicators for damage localization. We demonstrate the efficiency of the proposed schemes for damage analysis in a series of simulations using a three-mass swinger, in which the autoencoder can localize the damage using varying excitation locations. The covariances’ evaluation indicates that they are more potent than using the reconstruction error. Finally, experimental validation is conducted using vibration test data from a lattice tower called Leibniz University Structure for Monitoring (LUMO) under ambient excitation. For each damage pattern, high sensitivity towards local stiffness is achieved. Additionally, the damage position indicators exhibit a clear decreasing trend as the distance from the damage increases. The autoencoders presented here with non-standardized time series and covariance-based evaluation of residuals lead to increased robustness and sensitivity regarding damage localization.
KW - Autoencoder
KW - Data-driven model
KW - PCA
KW - Structural health monitoring
KW - Unsupervised damage localization
UR - http://www.scopus.com/inward/record.url?scp=85183452721&partnerID=8YFLogxK
U2 - 10.1016/j.engstruct.2024.117570
DO - 10.1016/j.engstruct.2024.117570
M3 - Article
AN - SCOPUS:85183452721
VL - 303
JO - Engineering structures
JF - Engineering structures
SN - 0141-0296
M1 - 117570
ER -