A probabilistic bond strength model for corroded reinforced concrete based on weighted averaging of non-fine-tuned machine learning models

Abstract

This paper develops an innovative probabilistic predictive model for bond strength of corroded reinforced concrete based on the weighted averaging of non-fine-tuned machine learning (ML) models. The model performs a weighted average of several ML-based predictive models using the Bayesian inference, so the model performance is improved and the time-consuming process of finding the optimized hyper-parameters can be avoided. Moreover, the model can also obtain the probabilistic information of the prediction. To build up the model, a database containing 384 datasets of relative bond strength for reinforced concrete was collected from the literature, which has nine input and one output variables. Four ML models without fine tuning are initially trained and then used to generate the final predictive model. The proposed model is compared with the four well-tuned ML models and five well-established empirical models. It is found that the predicted performance (i.e., accuracy and discrepancy) of the model is excellent and superior to those of the non-fine-tuned ML models and the empirical models, while can achieve comparable performance as the fine-tuned ML models. Furthermore, the mean, variance and probability distributions of the relative bond strength can all be attained by the model, which are more reasonable and accurate than the deterministic ML and empirical models.