Big Data Analytics in Structural Health Monitoring
This dissertation investigates methods to implement big data analytics in structural health monitoring. Four types of activities are considered: (1) data processing; (2) structural damage diagnosis and prognosis with uncertainty quantification; (3) high-dimensional model parameter calibration; and (4) surrogate model training. First, a methodology is developed to handle the various steps of data processing in structural health monitoring. MapReduce implementation is proposed to process sensor data of high volume, high velocity, and high variety. Then, techniques to parallelize structural damage diagnosis and prognosis with uncertainty quantification are developed. Both forward and inverse problems in uncertainty quantification are investigated with this efficient computational approach. Bayesian methods for the inverse problem of diagnosis, and numerical integration techniques such as Markov chain Monte Carlo (MCMC) simulation and Particle Filter (PF) are parallelized via MapReduce. Thirdly, high-dimensional model parameters calibration is performed efficiently using a three-level parallelization, in which spatial and temporal correlation is handled. Finally, parallelization of surrogate model training is developed, considering both response surrogates and distribution surrogates. Among distribution surrogates, a Gaussian mixture model is able to give analytical solutions for prediction and inference, which greatly reduces the cost of calibration of a high-dimensional model with high-volume data.