Structural Health Monitoring using Unsupervised Learning Methods with a view on Post-Earthquake Damage Detection
MetadataShow full item record
In a world of aging infrastructure, structural health monitoring (SHM) emerged as a major step towards resilient and sustainable societies. As we live in the information age, the advancements in machine learning and sensor technology have made SHM a more attractive damage detection method than the traditional non-destructive testing methods. Data-driven SHM requires a statistical learning model which could be supervised or unsupervised. Unsupervised learning only requires data from the undamaged structure during the training process. However, there is a lack of robust nonparametric density estimation to be used for the training model. In this thesis, a new unsupervised learning approach, multivariate Kernel Density Maximum Entropy method, is introduced in addition to examining two other approaches based on outlier analysis and kernel density estimation. Concepts and algorithms will be provided for each approach. Additionally, four types of damage-sensitive features are selected to be examined and compared in terms of provided advantages and disadvantages for each feature type. Two case studies, a numerical three-story three-dimensional reinforced concrete frame and a shake-table test of a three-story reinforced concrete frame with masonry infill, are investigated by applying the previously described unsupervised learning methods. In each case study, comparisons are conducted between results obtained by using different damage-sensitive features or different unsupervised learning approach. Conclusions and recommendations for future studies are provided at the end of this thesis.