Unveiling out-of-distribution data for reliable structural damage assessment in earthquake emergency situations

Abstract

To ensure accurate and effective emergency responses following an earthquake, one must promptly and accurately assess damage to structures with minimal manual effort. An advanced approach for evaluating structural damage is to use machine learning techniques for automated tagging or damage classification. However, most previous studies evaluated models using simple and unrealistic datasets limited to in-distribution (ID) datasets for training and testing. Although this approach may yield satisfactory results within the confines of the training dataset, it is not justifiable for real-time scenarios, in which the testing dataset may differ. Hence, a novel methodology is proposed herein that focuses on the detection of out-of-distribution (OOD) data. By subjecting a network to outliers, the model can effectively identify data outside the domain of the training dataset. A custom loss function is adopted, where both the cross-entropy loss from the ID training dataset and the log loss from the outlier dataset are incorporated. The effectiveness of this approach is demonstrated through its application in the post-event rapid damage assessment of bridges and shear walls subjected to seismic loading. A single network is employed to classify ID data into their respective classes, whereas the OOD data are shown to belong to the OOD class. The results highlight the significant accuracy achieved in the simultaneous prediction of ID and OOD data. By incorporating the detection of OOD data, this study enriches the methods for improving the reliability and accuracy of structural damage assessments in earthquake emergency responses, thus enabling more informed decision-making.