Data-driven stochastic model updating and damage detection with deep generative model

Is there a calibration algorithm beyond the dominant Bayesian sampling approach and sensitivity-based optimisation in model updating? Can a neural network serve not only as a surrogate model but also possess its own calibration capacity, independent of the Bayesian or optimisation framework? This work aims to address these questions by developing a unique data-driven approach for stochastic model updating and damage detection. Among a variety of models in deep learning, the class of deep generative model shares a similar objective, to estimate an unknown or intractable probability distribution from a small number of samples, with model updating. As a powerful flow-based deep generative model, a recently developed conditional Invertible Neural Network (cINN) architecture has been adopted in the task of model updating. Unlike the conventional approaches that employ the neural networks solely as a forward surrogate, the cINN-based model updating is a framework that performs as a bidirectional network where the forward training and inverse calibration are integrated into a uniform structure. The cINN consists of two parts known as the conditional network and the invertible neural network (INN). Both networks are trained jointly in the forward direction and can operate inversely to offer rapid and accurate predictions by given observation data. The application of the cINN provides a more efficient and direct manner to solve model updating problems without calculating the likelihood function in Bayesian inference. The cINN is embedded into a multilevel stochastic updating framework. Rather than directly calibrating physical parameters, this multilevel framework focuses on their statistical moments, e.g. mean and variance, referred to as hyperparameters. The hyperparameters are then utilised to determine the probability of damage (PoD), which provides a confidence level about the structural condition, facilitating stochastic damage detection. Two case studies are proposed to demonstrate the multilevel cINN-based stochastic updating and damage detection approach. The first involves a 3-degree-of-freedom spring-mass simulation model, while the second case study employs an experimental rig testcase with practical measurements, each under various damage scenarios.