Out-of-Distribution-Aware Structural Damage Classification via Hybrid Variational Autoencoder and Gradient Boosting

Structural damage classification is vital for the safety and longevity of critical infrastructure. Machine learning models, though effective in predicting structural behavior, often struggle with out-of-distribution data that differ from their training sets. While out-of-distribution detection has a...

Celý popis

Uloženo v:
Podrobná bibliografie
Vydáno v:Computers & structures Ročník 317; s. 107905
Hlavní autoři: Prayogo, Handy, Yang, I-Tung
Médium: Journal Article
Jazyk:angličtina
Vydáno: Elsevier Ltd 01.10.2025
Témata:
Gradient boosting
Machine learning
Out-of-distribution
Structural damage classification
Tabular data
Out-of-distribution
Structural damage classification
Tabular data
Machine learning
Gradient boosting
ISSN:0045-7949
On-line přístup:Získat plný text
Tagy: Přidat tag
Žádné tagy, Buďte první, kdo vytvoří štítek k tomuto záznamu!
Popis
Shrnutí:Structural damage classification is vital for the safety and longevity of critical infrastructure. Machine learning models, though effective in predicting structural behavior, often struggle with out-of-distribution data that differ from their training sets. While out-of-distribution detection has advanced for image and text data, progress remains limited for tabular data, the most common format representing critical structural information. This paper introduces a novel approach called Out-of-Distribution-Aware Hierarchical Gradient Boosting to address out-of-distribution challenges in tabular data within civil engineering applications. The method generates synthetic out-of-distribution samples, creating realistic samples that resemble in-distribution training data. The framework employs a two-stage hierarchical gradient boosting model: the first stage detects whether input data are in-distribution, and the second stage classifies structural damage within these samples. By integrating synthetic data generation with hierarchical modeling, the framework enhances the ability to detect out-of-distribution instances while maintaining classification accuracy. Gradient boosting algorithms including LightGBM, XGBoost, and CatBoost are used to balance robust detection and precise classification. Experimental results on the shear wall and bridge tagging datasets demonstrate that the proposed framework outperforms state-of-the-art multi-layer perceptron-based out-of-distribution detectors. These findings highlight a promising approach to enhancing infrastructure safety and performance specifically designed for tabular data.
ISSN:0045-7949
DOI:10.1016/j.compstruc.2025.107905