DescriptionThe traditional practice for condition evaluation of concrete bridge decks using GPR is limited to evaluating the upper section above the top reinforcement mat. As such, it does not provide any useful information about the condition of concrete below. In an attempt to expand the GPR evaluation zone beyond the top rebar, this research focuses on the development of a machine learning algorithm for the full depth condition assessment. Two learning algorithms were developed: (i) an algorithm based on numerical data, which was then applied to the experimental data from a GPR survey of a validation slab, and (ii) an algorithm based on a dataset comprised entirely of experimental data, which was later validated using a bridge and validation slab GPR survey data.
For the first algorithm, a database was developed through a series of two-dimensional numerical simulations of GPR surveys of slabs with variable influential or characteristic parameters. The slab was divided into three separate yet interconnected longitudinal layers. The quality of concrete was characterized by two electromagnetic properties – permittivity and conductivity, which were varied for each layer. Using the electromagnetic properties as characteristic parameters, six concrete conditions from good to critical were simulated. A machine learning technique, called gradient boosting, was used to predict the layers’ condition.
Gradient boosting was also used to analyze a dataset compiled through GPR surveys of four concrete bridge decks to predict the deck condition. Two independent prediction modules were developed: Module 1 to predict the condition above the top rebars, and Module 2 to predict the condition of concrete between the top and bottom rebars. A laboratory validation test slab and a fifth concrete highway bridge deck were surveyed using the same GPR system, and the data were used to validate the learning algorithm. The implementation of the proposed method in the validation phase showed that using machine learning and a vast library of GPR data, it is possible to avoid the arbitrary 90th percentile depth correction for new bridges without compromising the ability to assess the deck condition accurately. It was additionally demonstrated that it is possible for GPR to assess the condition of the deck beyond the top reinforcing mat.
To develop a more effective learning algorithm based on numerical simulations, it is recommended that a more extensive dataset is generated. It is also imperative to calibrate the numerical data with experimental data through laboratory testing in which the electromagnetic properties of concrete can be controlled at various depths.