Details
| Original language | English |
|---|---|
| Pages (from-to) | 637-660 |
| Number of pages | 24 |
| Journal | Structural health monitoring |
| Volume | 20 |
| Issue number | 2 |
| Early online date | 13 Mar 2020 |
| Publication status | Published - 13 Mar 2020 |
Abstract
This article proposes the deployment of adaptive boosting (AdaBoost) for combining damage feature decisions and improving the detection accuracy of structural health monitoring algorithms. In structural health monitoring applications, damage-sensitive features are combined with classifiers to define decision boundaries and provide information about the structural state. Boosting algorithms combine multiple classifiers aiming at the improvement of their performance. In this study, AdaBoost is deployed on the realizations of a modular structural health monitoring framework, which consists of three tiers: data normalization based on environmental and operational conditions; extraction of damage features, also referred to as condition parameters; and hypothesis testing. Each condition parameter–hypothesis testing pair composes a classifier which is used in AdaBoost as a weak classifier. The integration of AdaBoost with the structural health monitoring framework is validated using experimental data of a 3-kW wind turbine located at the Los Alamos National Laboratory and data generated from a mechanical model of the same structure. The AdaBoost classifier is evaluated with respect to the error rate as well as the true positive and false positive rates, which are typically used in receiver operating characteristic curves. The AdaBoost classifier outperforms the framework classifiers in many cases, improving drastically the detection performance. However, it is shown that the boosting performance depends on the relative location of the condition parameter values on the condition parameter space. The overlaps between the condition parameter values to be combined are quantified using the Bhattacharyya coefficient, which provides a metric for assessing the boosting potential. Finally, omitting condition parameter values corresponding to specific environmental and operational conditions from the boosting process is proposed for obtaining optimum boosting results.
Keywords
- AdaBoost, boosting, damage detection, structural health monitoring, wind turbines
ASJC Scopus subject areas
- Biochemistry, Genetics and Molecular Biology(all)
- Biophysics
- Engineering(all)
- Mechanical Engineering
Cite this
- Standard
- Harvard
- Apa
- Vancouver
- BibTeX
- RIS
In: Structural health monitoring, Vol. 20, No. 2, 13.03.2020, p. 637-660.
Research output: Contribution to journal › Article › Research › peer review
}
TY - JOUR
T1 - Combination of damage feature decisions with adaptive boosting for improving the detection performance of a structural health monitoring framework
T2 - Validation on an operating wind turbine
AU - Tsiapoki, Stavroula
AU - Bahrami, Omid
AU - Häckell, Moritz W.
AU - Lynch, Jerome P.
AU - Rolfes, Raimund
N1 - Funding information: The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Science Foundation under Grants CMMI-1362513 and CMMI-1436631.
PY - 2020/3/13
Y1 - 2020/3/13
N2 - This article proposes the deployment of adaptive boosting (AdaBoost) for combining damage feature decisions and improving the detection accuracy of structural health monitoring algorithms. In structural health monitoring applications, damage-sensitive features are combined with classifiers to define decision boundaries and provide information about the structural state. Boosting algorithms combine multiple classifiers aiming at the improvement of their performance. In this study, AdaBoost is deployed on the realizations of a modular structural health monitoring framework, which consists of three tiers: data normalization based on environmental and operational conditions; extraction of damage features, also referred to as condition parameters; and hypothesis testing. Each condition parameter–hypothesis testing pair composes a classifier which is used in AdaBoost as a weak classifier. The integration of AdaBoost with the structural health monitoring framework is validated using experimental data of a 3-kW wind turbine located at the Los Alamos National Laboratory and data generated from a mechanical model of the same structure. The AdaBoost classifier is evaluated with respect to the error rate as well as the true positive and false positive rates, which are typically used in receiver operating characteristic curves. The AdaBoost classifier outperforms the framework classifiers in many cases, improving drastically the detection performance. However, it is shown that the boosting performance depends on the relative location of the condition parameter values on the condition parameter space. The overlaps between the condition parameter values to be combined are quantified using the Bhattacharyya coefficient, which provides a metric for assessing the boosting potential. Finally, omitting condition parameter values corresponding to specific environmental and operational conditions from the boosting process is proposed for obtaining optimum boosting results.
AB - This article proposes the deployment of adaptive boosting (AdaBoost) for combining damage feature decisions and improving the detection accuracy of structural health monitoring algorithms. In structural health monitoring applications, damage-sensitive features are combined with classifiers to define decision boundaries and provide information about the structural state. Boosting algorithms combine multiple classifiers aiming at the improvement of their performance. In this study, AdaBoost is deployed on the realizations of a modular structural health monitoring framework, which consists of three tiers: data normalization based on environmental and operational conditions; extraction of damage features, also referred to as condition parameters; and hypothesis testing. Each condition parameter–hypothesis testing pair composes a classifier which is used in AdaBoost as a weak classifier. The integration of AdaBoost with the structural health monitoring framework is validated using experimental data of a 3-kW wind turbine located at the Los Alamos National Laboratory and data generated from a mechanical model of the same structure. The AdaBoost classifier is evaluated with respect to the error rate as well as the true positive and false positive rates, which are typically used in receiver operating characteristic curves. The AdaBoost classifier outperforms the framework classifiers in many cases, improving drastically the detection performance. However, it is shown that the boosting performance depends on the relative location of the condition parameter values on the condition parameter space. The overlaps between the condition parameter values to be combined are quantified using the Bhattacharyya coefficient, which provides a metric for assessing the boosting potential. Finally, omitting condition parameter values corresponding to specific environmental and operational conditions from the boosting process is proposed for obtaining optimum boosting results.
KW - AdaBoost
KW - boosting
KW - damage detection
KW - structural health monitoring
KW - wind turbines
UR - http://www.scopus.com/inward/record.url?scp=85082106889&partnerID=8YFLogxK
U2 - 10.1177/1475921720909379
DO - 10.1177/1475921720909379
M3 - Article
AN - SCOPUS:85082106889
VL - 20
SP - 637
EP - 660
JO - Structural health monitoring
JF - Structural health monitoring
SN - 1475-9217
IS - 2
ER -