Fully decouple convolutional network for damage detection of rebars in RC beams

Abstract

This paper proposes a fully decouple convolutional neural network equipped with an anchor-free mechanism for the non-destructive identification of structures with internal damages. Two 3D cylinder IoU optimization and evaluation algorithms are designed to improve model training effect and to estimate model performance, respectively. The generation and preparation of the datasets is exceedingly time consuming and challenging, to overcome these difficulties, datasets of random damages of rebars in RC beams are established based on the secondary development of ABAQUS. The mapping relationship of the model takes the displacement contour image of damaged structures as the input and damage parameters of rebars as the output. The image feature extraction is performed to capture damage features in surface displacement information of the damaged structures based on proposed model, then, the structure internal damage identification is realized through multi-regression on damage parameters of rebars. To verify the efficiency of the proposed method, model comparison and numerical verification is utilized to evaluate the proposed model. The results show that the presented model has high accuracy and is superior to traditional coupled networks. The method can also be extended to the internal damage detection of other components, providing a new idea for the internal non-destructive health monitoring of prefabricated structures.