Model-Class Selection Using Clustering and Classification for Structural Identification and Prediction

Structural identification using physics-based models and subsequent prediction have much potential to enhance civil infrastructure asset-management decision-making. Interpreting monitoring information in the presence of multiple uncertainty sources and systematic bias using a physics-based model is a computationally expensive task. The computational cost of this task is exponentially proportional to the number of model parameters updated using monitoring data. In this paper, a novel model-class selection method is proposed to obtain computationally optimal and identifiable model classes. Unlike traditional sensitivity methods for model-class selection, in the proposed method, model responses at sensor locations are clustered to identify underlying trends in model response datasets. K-means clustering is used to determine relevant clusters in the data. Cluster indices are then used as labels for classification. Support-vector machine classification using forward variable selection with sequential search is used to select model parameters that help classify trends in data. The result of the sequential search is a trade-off curve comparing classification error with number of parameters in the model class. This curve helps select a practical and near-optimal model class. The model-class selection method proposed in this paper is compared with linear regression-based sensitivity analysis using a full-scale bridge. Identification with model classes obtained using both methods for two sensor configurations suggests that the model-based clustering method helps select an identifiable and computationally efficient model class. The minimum remaining fatigue life of the bridge predicted using the updated model classes is 720 years and this represents fatigue-life extension of 10 times compared with design predictions prior to measurements. This approach provides good support for asset managers when they interpret measurement data.