Deep Learning Based Rebar Signal Suppression And Defect Recognition Of Ground Penetrating Radar Image

Reinforced concrete is the most widely used building material in infrastructure such as bridges,tunnels,and dams.Long-term reinforced concrete structures are prone to internal defects such as cracks,voids,holes,and incompactness,which seriously threaten the healthy operation of large-scale infrastructure.Ground penetrating radar(GPR)has been widely used in rebar detection,rebar depth estimation,concrete defect identification and other fields due to its advantages of non-destructive detection,strong penetrating ability,high speed and high efficiency,etc.Due to the shielding effect of the rebar layer on the electromagnetic wave of the GPR,the signal of the rebar in the B-scan of the GPR seriously interferes the signal of the defect below it,which makes it extremely difficult to identify the defect under the rebar.Therefore,it has become an urgent need to study an accurate and efficient method for defect identification under rebar and to improve the accuracy of defect identification under rebar.The main research work and achievements of this paper are as follows:(1)Aiming at the inherent problem that the rebar signal seriously obscured the defect signal in the GPR image,which made it difficult to identify the defect under the rebar,a rebar signal suppression and defect signal reconstruction method based on deep learning was studied,and a model of rebar signal suppression based on supervised learning was established.The suppression model focused on the characteristics of the GPR image.The Residual-Inception module and the attention module were designed to extract rich features of different scales,guiding the model to pay attention to the weak disease signal,and solved the problem that the serious interference of the rebar signal made the defect signal difficult to accurately reconstruct.Comparative experiments and ablation experiments were carried out on the simulation data set and the noise-added simulation data set to test the ability of rebar signal suppression and accurate reconstruction of the defect.Finally,a sandbox test was carried out to test the effect of the suppression model on real data.(2)Aiming at solving the problem of missing data sets caused by the difficulty in establishing the corresponding relationship between the GPR image of the defect under the rebar and the GPR image of the defect without the rebar in actual projects,a method of rebar signal suppression and defect reconstruction based on unsupervised learning was proposed,and a rebar signal suppression model based on generative adversarial network was constructed.Relying on the task requirements of accurately reconstructing the defect signal after rebar signal suppression,feature encoders with different functions were added to the model.Contrast loss and similarity loss were introduced to solve the problem of inaccurate position and shape of the reconstructed defect signal.Through the comparison experiment and ablation experiment,the suppression effect of the model was tested,and the applicability of the method on the actual GPR data was verified by the sandbox model test.(3)The GPR data collected in actual engineering had various scenarios(different data forms,environmental noises,frequencies,etc.),therefore,it is difficult to establish a large number of labeled data sets for each scenario,which results in low accuracy rate of defect classification with rebar data in different scenarios.To solve this problem,this paper studied an anomaly classification method based on meta-learning for data including rebar signal,and improved the classification accuracy of classification models in different scenarios by using a very small amount of data in different scenarios as standard reference samples.Through comparative experiments on data from different scenarios,the model performance was tested.Finally,a series of methods were established for identifying defect under rebar including anomaly classification,rebar suppression and defect identification,and the accurate identification of defect under rebar was realized as well.