Predicting the microbiologically induced concrete corrosion in sewer based on XGBoost algorithm

Abstract

Microbiologically induced concrete corrosion (MICC) is a major reason for sewer pipe replacement and rehabilitation. Predicting the depth of concrete corrosion can help make better decisions on the inspection and rehabilitation of pipes. Due to the limited data available, the models using deep learning methods for predicting MICC are prone to overfitting. This study develops an XGBoost-based MICC model with the benefits of hyperparametric auto-optimization. For this purpose, the factors affecting MICC were sorted out to be used as a guide for selecting attributes when building the database. 379 datasets relevant to the corrosion loss of the origin Portland cement-based materials in sewers were collected from literature to establish a database. It is then randomly partitioned into 8:2 training and test datasets. The hyperparameters were adjusted by performing Bayesian optimization for 200 iterations. On the training and test sets, the R2 scores of the trained model are 0.87 and 0.85, respectively. In addition, an ongoing field trail in a sewer well is introduced to test the generalization ability of the model. By deploying the model on the field inspection data, the predicted corrosion loss after 169 days is 1.6 mm, which ranges between 1.5 and 3.5 mm measured. The developed MICC model is anticipated to perform well when fed future inspection data that is richer and more abundant.

Keywords

• Sewer pipe
• Microbiologically induced concrete corrosion
• Corrosion model
• XGBoost
• Bayesian optimization