Recognition of the Distress in Concrete Pavement Using Deep Learning Based on GPR Image

Abstract

Road testing results are the basement of road maintenance decision-making. However traditional drilling core damage the structural integrity of the pavement, and work period is also disadvantage in the maintenance of the road. Ground penetrating radar (GPR) which is non-destructive and high efficiency, has outstanding advantages and become the most important way to check the internal distress of the concrete pavement, but the subject remains a challenge in the acquired radar signal interpretation, especially under noisy interference the traditional methods is difficult to recognize the distress accurately. Based on the current research status of the concrete pavement inspection, we proposed GPR pavement distress automatic recognition technology based on the deep learning. Through simulation technology, the various types of distress in the inspection of pavement concrete using GPR was simulated such as void, disengaging and loose. Based on GPR data collection, the features of the concrete pavement is analyzed. The deep learning models was established for recognizing distress concrete pavement structure. The establishment of the sample was from simulation results of concrete pavement, then the datasets were built. Recurrent neural networks (RNN), convolution neural networks (CNN) and deep belief networks (DBN) were compared in the study, and the advantages of automatic deep belief networks is shown in recognizing distress. Finally, stacking automatic coding machine was applied to practical engineering. Based on experiment, the proposed method accurately recognize the concrete pavement distress, and the type of the distress also classified.