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

Internal crack recognition of reinforce concrete structure based on array ultrasound and feature fusion neural network

在工程实践和已有研究中,对于混凝土结构内部病害的无损检测,超声层析图像通常需要先验知识进行人工定性判读,较少应用于精确的量化检测。为此,提出一种基于阵列超声和特征融合神经网络的深度学习方法,用于钢筋混凝土结构内部裂缝的像素级无损检测。制作了预置内部裂缝的钢筋混凝土构件,并使用低频剪切超声波和换能器阵列扫查,获得阵列超声B扫描图像并构建数据集。建立了具有编码器-解码器架构的深度神经网络,并采用特征融合策略和残差模块优化该模型,使其与B扫描图像的语义结构更加兼容。通过配准,将独立局部预测图组合为断面的全局表征,以展示构件长度方向整个断面的裂缝位置、分布等全局信息。结果表明:训练、验证及测试集的F分数均高于70%,提出的特征融合深度神经网络可识别宽度低至1 mm的裂缝,裂缝长度量化的平均绝对百分误差为6.22%,证实了所提方法的有效性。

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 1mm 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.