Real-time unsupervised structural damage detection using cepstral features and principal component analysis

In structural health monitoring, the rapid detection and classification of structural damage are critical for ensuring the safety and longevity of civil infrastructures. In this context, supervised learning techniques that require extensive labeled undamaged and damaged data are often impractical, given the lack of observation of structural systems under different and exhaustive failure modes. This article introduces a novel unsupervised learning approach that utilizes cepstral coefficients (CCs) as damage sensitive features (DSFs). The aim is to develop a robust, real-time algorithm capable of detecting and classifying structural damage without needing pre-labeled data. This approach is particularly valuable in situations where immediate structural integrity assessment is crucial. Moreover, in addition to enabling binary assessment between damaged and undamaged conditions, such an algorithm can also recognize different levels of damage. The proposed strategy relies on CCs, DSFs extracted from the structure’s vibration response to detect variations in structural behavior. Recent studies have demonstrated the effectiveness of such coefficients in capturing the structural properties of dynamic systems well, and their ease of extraction makes them particularly suitable for real-time monitoring of bridges, buildings, and other infrastructure systems. The proposed methodology consists of an online novelty detection strategy based on principal component analysis of the identified coefficients to enhance differentiation between coefficients extracted from different health conditions. The squared Mahalanobis distance is used as a discrimination threshold tool for outlier detection. Results are presented using real-world data extracted from the Z24 bridge to assess the strategy’s effectiveness. The results demonstrate the algorithm’s performance in differentiating between damaged and undamaged data and its capability to accurately classify incoming data associated with various levels of damage.