ABSTRACT
Corrosion-induced crack widths provide valuable information about the corrosion states of steel reinforcements embedded in reinforced concrete (RC) structures. The uncertain relationship between corrosion-induced cracking and steel weight loss (SWL) necessitates a probabilistic approach to estimate the spatial distribution of SWL, which is essential for evaluating the capacity loss of corroded RC structures. This paper presents a Bayesian framework for inferring SWL distributions based on observed corrosion-induced crack widths. A Karhunen-Loève transformation reduces the inference dimensions by extracting principal features of the SWL distribution. The forward model utilizes a data-driven sequence-to-sequence approach, which combines a nonlinear convolution kernel for input encoding with a sparse polynomial chaos expansion for decoding. The framework effectively estimates the posterior distribution of SWL using the Hamiltonian Markov chain Monte Carlo sampler, which leverages gradient information to enhance sampling efficiency. The numerical application of the proposed method demonstrates its robustness, with 95% confidence intervals successfully encompassing most SWL observations.
