Probabilistic prediction model for critical chloride concentration of reinforcement corrosion based on improved Gaussian process regression

To describe the probabilistic characteristics of critical chloride concentration (CCC) and calibrate the computational accuracy of traditional prediction models, a probabilistic prediction model of CCC for reinforcement corrosion in concrete has been developed based on an improved Gaussian process regression (GPR) model. Firstly, a new hybrid kernel function for the GPR model is proposed by combining the radial basis function (RBF) with the rational quartic kernel (RQK) according to the sum of single kernel functions and automatic relevance determination function. Then the hyperparameters of the improved GPR model are determined based on the Bayesian theory and the maximum likelihood estimation. Subsequently, a probabilistic prediction model of CCC for reinforcement corrosion in concrete is developed based on the improved GPR model and a total of 591 sets of experimental data. Finally, the accuracy and applicability of the proposed probabilistic prediction (PPP) model is validated by comparing with traditional kernel functions, machine learning models and empirical theoretical models. The results show that the PPP model based on the new hybrid kernel function has high accuracy and generalisation ability. The PPP model provides efficient way to describe the probabilistic characteristics of CCC for reinforcement corrosion in concrete and to calibrate the computational accuracy of traditional prediction models based on probability density function and confidence interval.