Probabilistic assessment of structural condition through clustering-based multiple FE model update

Abstract

This study proposes a new procedure for probabilistic assessment of structural condition through updating multiple FE models. Sets of measurement data are constructed in order to incorporate measurement uncertainties into the model updating process. Multiple updated FE models are obtained by performing optimization for each constructed set of measurement data. Each updated FE model is regarded as the most probable one that represents the state of the concerned structure for the given measurements. As a result, variability of possible structural conditions caused by the uncertainties in measurement data can be inferred from the statistical distribution of multiple FE models. In addition, machine learning techniques are employed for the purpose of probabilistic description and classification of the multiple updated FE models. The Principal Component Analysis (PCA) method is utilized to better understand the distribution feature of multiple FE models by transforming FE models onto principal subspace. The sensitivity of structural parameters to the uncertainty of data is also investigated by using the principal components. Moreover, overall variance characteristics can be described more efficiently using a subset of the principal components. The transformed models are further classified based on their similarity by using the K-means method. The probabilistic features of the FE models are then identified by fitting a Gaussian mixture model to the distribution. Finally, the statistical features of structural condition are obtained by using the identified classes of the updated FE models. The proposed procedure is demonstrated by the numerical example of the Yeondae Bridge, a 4-span continuous steel-box girder bridge in South Korea. The distribution of the rating factors is evaluated using the updated FE models. The probability of the bridge failure is also estimated by structural reliability analysis utilizing the identified Gaussian mixture models. The results show that the clustering-based model selection procedure can reduce the possibility of inaccurate condition assessment caused by a high level of measurement error, thus can provide more consistent rating factor and reliability index.