One-class machine learning approach for localized damage detection

With the advancement in computing power and sensing technology in the last decade, smart monitoring and decision-making approaches are becoming more feasible and replacing the traditional structural health monitoring (SHM) techniques. In this paper, an unsupervised machine learning approach, namely one-class support vector machine (OC-SVM) approach augmented with a localization metric is proposed for damage identification purposes. The auto-regressive (AR) parameters estimated using acceleration measurements collected from the sensors are utilized as damage sensitive features to train the OC-SVM and damage detection stage is accomplished in a decentralized manner where data acquired with each sensor are processed locally. The statistical measures; Itakura distance, Mahalanobis distance, and Fisher criterion are employed as possible means to localize damage once damage is detected by the trained machine. The method is evaluated through numerical simulations on a five-story moment resisting frame with several damage scenarios including loss of member and connection rigidities. The assessment of the methodology is completed through experimental data collected on a one-story one-bay reinforced concrete frame subjected to progressive damage. The proposed methodology shows promise towards implementation as an automated damage assessment tool that can be employed for frame type structures.