基于阵列超声和特征融合神经网络的钢筋混凝土结构内部裂缝检测OA北大核心CSTPCD

In existing studies and practical nondestructive testing applications,ultrasonic tomography images were usually utilized for manual qualitative interpretation but hardly used for accurate quantitative detection purposes of internal defects for reinforced concrete(RC)structures.To this end,a deep learning method based on array ultrasound and feature fusion neural network was proposed in this study for pixel-wise nondestructive recognition of internal cracks in RC structures.RC components with preset artificial internal cracks were manufactured.Array ultrasonic B-scan images were then acquired by testing the RC components with shear-wave low-frequency transducer array,and the dataset was setup.A deep neural network with the basic encoder-decoder architecture was developed,which was optimized by feature fusion strategy and residual modules to improve the compatibility with the semantic structure of ultrasonic B-scans.Moreover,individual local predicted images were combined with global representations by registration to indicate global information such as crack location and distribution of the entire section.The results indicate that F-scores of the training,validation,and testing sets are higher than 70％.The cracks as small as 1 mm in width can be recognized by the proposed feature fusion neural network,and the mean absolute percentage error of quantified crack length is 6.22％,substantiating the effectiveness of the proposed method.