Early warning of structural damage via manifold learning-aided data clustering and non-parametric probabilistic anomaly detection

Unsupervised learning is an effective and practical methodology for structural health monitoring when the preparation of labeled training data regarding damaged states is intractable, unnecessary, and expensive. Despite several studies on this field, some challenging issues need further evaluations. The main objective of this research is to overcome the challenges concerning different variability patterns in unlabeled vibration data caused by single and multiple environmental and/or operational variations, non-generality of unsupervised learners in handling training data of different sizes, and impacts of false positive and false negative errors on early damage warning. A novel integrated unsupervised learning method is proposed that emanates from manifold learning-aided data clustering and non-parametric probabilistic anomaly detection. Data clustering is based on regularized Gaussian mixture modeling supported by nearest neighbor graphs, for which training data can be modeled on a manifold structure. The primary purpose of this step is to generate local subsets of the entire training data in an effort to minimize the effects of environmental and/or operational variations. A multi-fidelity hyperparameter optimization is also designed to set the main hyperparameters of the proposed clustering algorithm, namely the number of components (clusters) and a regularization value. Using the aforementioned local subsets, a non-parametric probabilistic anomaly detector is developed from a reverse Gaussian mixture function to compute anomaly scores for early damage warning. Modal frequencies regarding two large-scale bridges are used to validate the proposed method and compare it with some state-of-the-art techniques. Results confirm the effectiveness and reliability of this method with negligible errors under different environmental variability.