Unsupervised feature selection-based technique for locating structural deterioration: a multi-domain approach

Resumo

Structural monitoring methods have been extensively researched in recent years due to developments

in Artificial Intelligence (AI) technology. In this regard, the purpose of this work is to offer an automated data-
driven approach for deterioration localization based on the extraction of features from raw vibration data utilizing

domain knowledge and a filtering procedure. To diversify information retrieval, feature extraction is conducted
concurrently in temporal, frequency, and quefrency domains. This filtering process is known as feature selection
(FS) and is used to reduce redundancies and raise the relevance of the feature set by removing a subset based on a
predefined criterion. The key idea is that the proposed approach may be tuned to the structure while offering
generality for whatever shape, material, or excitation it comes across. The deterioration index is calculated via

outlier analysis referenced by the structure's healthy condition. The technique was successfully tested in a full-
scale bridge, demonstrating a performance that is encouraging for real-world monitoring scenarios.