An Unsupervised Damage Detection Strategy for Recognizing Unseen Health Conditions in Monitoring Bridges

The majority of damage assessment strategies for civil and mechanical systems primarily focus on the binary classification between healthy or damaged conditions, often lacking the ability to differentiate between different levels or types of damage and recognize unseen conditions. This paper introduces a novel unsupervised learning approach that adopts a probabilistic framework for damage classification able to address these critical needs. The proposed method aims to differentiate between different health conditions utilizing strategic features derived directly from the vibrational response of the system. Classification is performed by first, learning the features distribution from the healthy state of the system, considered as reference state; then the likelihood of new incoming data from an unknown state is evaluated to determine to which class the unknown observation belongs. When the likelihood falls below a predefined threshold, a new damage class is created and learned, allowing for more detailed and accurate damage characterization. The proposed methodology combines Probabilistic Linear Discriminant Analysis (PLDA) with the use of cepstral features, adopted as damage sensitive features. These coefficients are traditionally used in speaker recognition and recent research studies have demonstrated that such features effectively capture the structural properties of dynamic systems, simplifying the analysis process while enhancing efficiency and effectiveness. This innovative approach to damage classification not only could serve as an early warning system for the system’s health but could also become a valuable tool for predictive maintenance operations in both civil structures and mechanical systems.