Data-driven Concrete Damage Diagnosis with Thermal Imaging and Vibration Testing
This dissertation work investigates data-driven approaches for concrete structures health monitoring by considering multiple monitoring techniques and quantifies the uncertainty in the diagnosis result of each method. First, the dissertation develops an interior damage diagnosis (detection, localization, and quantification) methodology with traditional image processing techniques using thermography data. Second, a two-stage damage diagnosis (detection and localization) methodology is proposed with innovative damage-sensitive features using time-series vibration data. The damage-sensitive features are based on singular value decomposition (SVD) and crest factor metric. Third, a feature automation and interior damage diagnosis methodology is developed using the deep convolutional neural network approach and transfer learning in problems with limited and small datasets. Finite element models are exploited to augment the dataset. A systematic parametric selection method for the finite element models is provided. All the proposed methodologies are tested and validated using datasets from experiments on concrete, which is a heterogeneous material. Specimens (damaged and undamaged) with different sizes (involving thin patio block samples and thick concrete block samples) are utilized during the experiments. The proposed methodologies provide promising performance with consistently high accuracy in diagnosis results on the datasets from various specimens. Within each of the methodologies, the reliability of the data-driven approach is assessed through uncertainty and robustness analyses using techniques such as global sensitivity analysis, Bayesian updating, Markov Chain Monte Carlo, and Monte Carlo dropout.