A platform for research: civil engineering, architecture and urbanism
A platform for research: civil engineering, architecture and urbanism
Structural Dynamics Feature Learning Using a Supervised Variational Autoencoder
https://orcid.org/0000-0002-0834-185X
https://orcid.org/0000-0002-1103-7699
https://orcid.org/0000-0002-6870-240X
In this study, we propose a novel approach for learning structural dynamics and discover useful features using a supervised variational autoencoder (SVAE) with the aim of executing classification tasks that feed into structural health monitoring (SHM) schemes for damage detection and characterization, as well as transfer learning. The SVAE is trained on dynamic response data, which can be drawn from either simulated or experimentally measured responses of monitored systems, and is charged with executing an auxiliary classification into categories (labels) on the basis of known system properties that may reflect stiffness, roughness, or other characteristics. This allows the model to learn a compact and expressive representation of the dynamical features of the structure generalizing across different property clusters. We evaluate the performance of the SVAE by comparing its ability to reconstruct unseen response measurements with that of the variational autoencoder (VAE) and conditional VAE (CVAE) variants. The VAE does not allow for conditioning on external variables, whereas the CVAE allows for conditioning on—typically—continuous parameters but requires these as inputs at runtime as well. In a first illustrative example, we show that the SVAE accurately captures the dynamics of the structure, conditioned on expected ranges of influencing parameters, and outperforms VAE and CVAE-based models in terms of reconstruction accuracy. We then illustrate the impact of our suggested method on more complex data sets. The second example demonstrates the efficacy of the approach on highly noisy field data derived from an instrumented bicycle used to traverse different types of roads. The SVAE scheme is shown to be highly capable in classifying the different types of road surfaces based on their pavement material or quality. In the final example, we apply the SVAE model on data related to the dynamics of a complex structure, namely, a wind turbine. We form a synthetic data set that models evolving delamination on a wind turbine blade, for which we show that the use of the SVAE allows for the identification of such damage using a low number of sensors. Therefore, this article places SVAE on the map as a salient candidate for dynamics and damage characterization tasks in the context of SHM.
Structural Dynamics Feature Learning Using a Supervised Variational Autoencoder
https://orcid.org/0000-0002-0834-185X
https://orcid.org/0000-0002-1103-7699
https://orcid.org/0000-0002-6870-240X
In this study, we propose a novel approach for learning structural dynamics and discover useful features using a supervised variational autoencoder (SVAE) with the aim of executing classification tasks that feed into structural health monitoring (SHM) schemes for damage detection and characterization, as well as transfer learning. The SVAE is trained on dynamic response data, which can be drawn from either simulated or experimentally measured responses of monitored systems, and is charged with executing an auxiliary classification into categories (labels) on the basis of known system properties that may reflect stiffness, roughness, or other characteristics. This allows the model to learn a compact and expressive representation of the dynamical features of the structure generalizing across different property clusters. We evaluate the performance of the SVAE by comparing its ability to reconstruct unseen response measurements with that of the variational autoencoder (VAE) and conditional VAE (CVAE) variants. The VAE does not allow for conditioning on external variables, whereas the CVAE allows for conditioning on—typically—continuous parameters but requires these as inputs at runtime as well. In a first illustrative example, we show that the SVAE accurately captures the dynamics of the structure, conditioned on expected ranges of influencing parameters, and outperforms VAE and CVAE-based models in terms of reconstruction accuracy. We then illustrate the impact of our suggested method on more complex data sets. The second example demonstrates the efficacy of the approach on highly noisy field data derived from an instrumented bicycle used to traverse different types of roads. The SVAE scheme is shown to be highly capable in classifying the different types of road surfaces based on their pavement material or quality. In the final example, we apply the SVAE model on data related to the dynamics of a complex structure, namely, a wind turbine. We form a synthetic data set that models evolving delamination on a wind turbine blade, for which we show that the use of the SVAE allows for the identification of such damage using a low number of sensors. Therefore, this article places SVAE on the map as a salient candidate for dynamics and damage characterization tasks in the context of SHM.
Structural Dynamics Feature Learning Using a Supervised Variational Autoencoder
https://orcid.org/0000-0002-0834-185X
https://orcid.org/0000-0002-1103-7699
https://orcid.org/0000-0002-6870-240X
In this study, we propose a novel approach for learning structural dynamics and discover useful features using a supervised variational autoencoder (SVAE) with the aim of executing classification tasks that feed into structural health monitoring (SHM) schemes for damage detection and characterization, as well as transfer learning. The SVAE is trained on dynamic response data, which can be drawn from either simulated or experimentally measured responses of monitored systems, and is charged with executing an auxiliary classification into categories (labels) on the basis of known system properties that may reflect stiffness, roughness, or other characteristics. This allows the model to learn a compact and expressive representation of the dynamical features of the structure generalizing across different property clusters. We evaluate the performance of the SVAE by comparing its ability to reconstruct unseen response measurements with that of the variational autoencoder (VAE) and conditional VAE (CVAE) variants. The VAE does not allow for conditioning on external variables, whereas the CVAE allows for conditioning on—typically—continuous parameters but requires these as inputs at runtime as well. In a first illustrative example, we show that the SVAE accurately captures the dynamics of the structure, conditioned on expected ranges of influencing parameters, and outperforms VAE and CVAE-based models in terms of reconstruction accuracy. We then illustrate the impact of our suggested method on more complex data sets. The second example demonstrates the efficacy of the approach on highly noisy field data derived from an instrumented bicycle used to traverse different types of roads. The SVAE scheme is shown to be highly capable in classifying the different types of road surfaces based on their pavement material or quality. In the final example, we apply the SVAE model on data related to the dynamics of a complex structure, namely, a wind turbine. We form a synthetic data set that models evolving delamination on a wind turbine blade, for which we show that the use of the SVAE allows for the identification of such damage using a low number of sensors. Therefore, this article places SVAE on the map as a salient candidate for dynamics and damage characterization tasks in the context of SHM.
Structural Dynamics Feature Learning Using a Supervised Variational Autoencoder
J. Eng. Mech.
Bacsa, Kiran (author) / Liu, Wei (author) / Abdallah, Imad (author) / Chatzi, Eleni (author)
2025-02-01
Article (Journal)
Electronic Resource
English
Optimization of force-field potential parameters using conditional variational autoencoder
| DOAJ | 2023
Generative design of RHS/SHS profiles using Conditional Variational Autoencoder (CVAE)
| TIBKAT | 2024