Pattern recognition algorithms for density estimation of asphalt pavement during compaction: a simulation study

This paper presents the application of artificial neural network (ANN) based pattern recognition to extract the density information of asphalt pavement from simulated ground penetrating radar (GPR) signals. This study is part of research efforts into the application of GPR to monitor asphalt pavement density during compaction. The main challenge is to eliminate the effect of roller-sprayed water on GPR signals during compaction and to extract density information accurately. A calibration of the excitation function was conducted to provide an accurate match between the simulated signal and the real signal. A modified electromagnetic mixing model was then used to calculate the dielectric constant of asphalt mixture with water. A large database of GPR responses was generated from pavement models having different air void contents and various surface moisture contents using finite-difference time-domain simulation. Feature extraction was performed to extract density-related features from the simulated GPR responses. Air void contents were divided into five classes representing different compaction statuses. An ANN-based pattern recognition system was trained using the extracted features as inputs and air void content classes as target outputs. Accuracy of the system was tested using test data set. Classification of air void contents using the developed algorithm is found to be highly accurate, which indicates effectiveness of this method to predict asphalt concrete density.