Bayesian Neural Networks vs. Model Updating for Inference of Miter Gate Damage

Abstract

Miter gates are critical components of inland waterways and their failure can have significant consequences. These gates must fully contact a wall or stress concentrations develop, leading to cracking and failure. To detect gaps in the contact, we compare two inference methods: Bayesian model updating using the transitional Markov chain Monte Carlo (TMCMC) method and neural network-based Bayesian inference. TMCMC is a widely used method, but inference is computationally expensive. In contrast, neural network-based Bayesian inference trains a neural network upfront and infers damage parameters using the computationally efficient neural network, dramatically reducing runtime as more datasets are gathered and parameters are updated. Through a case study data gathered at a specific miter gate, both methods are compared on their performance inferring boundary condition degradation using vision- based displacement data. Results show that neural network-based Bayesian inference achieves similar accuracy to TMCMC updating with a significant reduction in computational costs from hours to seconds. These computational advantages may aid in online and edge applications at miter gates and other civil structures. With regular updates of miter gate bearing contact damage, decision-makers can more effectively maintain miter gates.