CN-113159052-B - Method for identifying failure mode of flexural reinforced concrete simply supported beam based on deep learning

CN113159052BCN 113159052 BCN113159052 BCN 113159052BCN-113159052-B

Abstract

A method for identifying a failure mode of a bent reinforced concrete simply supported beam based on deep learning belongs to the field of failure mode identification of reinforced concrete members. The steps are as follows: establishing a reinforced concrete structure surface crack and peeling damage semantic segmentation data set; constructing a deep learning semantic segmentation full convolution network for identifying surface cracks and peeling damage pixels of the reinforced concrete structure; training and verifying a built deep learning semantic segmentation full convolution network; adopting a training and verifying deep learning semantic segmentation full convolution network model to identify cracks and peeling damage pixels in the surface image of the bent reinforced concrete simply supported beam; extracting the characteristic of damage of the normal section or the inclined section of the bent reinforced concrete simply supported beam in the identified image; and carrying out damage mode identification on the bent reinforced concrete simply supported beam according to the identified characteristics of the normal section or the inclined section. And expensive detection equipment is not needed, and the degree of automation of failure mode identification of the bent reinforced concrete simply supported beam is improved.

Inventors

LI SHENGYUAN
Ding beidou
ZHANG YINGYING
LU DONGYUAN

Assignees

中国矿业大学
中国矿业大学

Dates

Publication Date: 20231117
Application Date: 20210427
Priority Date: 20210427

Claims (3)

1. A method for identifying a failure mode of a bent reinforced concrete simply supported beam based on deep learning is characterized by comprising the following steps: the method comprises the following steps: s1, establishing a reinforced concrete structure surface crack and peeling damage semantic segmentation data set; s2, constructing a deep learning semantic segmentation full convolution network for identifying surface cracks and peeling damage pixels of the reinforced concrete structure; s3, training and verifying the deep learning semantic segmentation full convolution network constructed in the S2 by adopting the data set constructed in the S1; s4, performing crack and peeling damage pixel identification in the surface image of the bent reinforced concrete simply supported beam by adopting the deep learning semantic segmentation full convolution network model trained and verified in the S3; s5, extracting the damaged characteristics of the positive section or the inclined section of the bent reinforced concrete simply supported beam in the image identified in the S4; s6, identifying the damage mode of the bent reinforced concrete simply supported beam according to the damage characteristics of the normal section or the inclined section extracted in the S5; the concrete structure surface microcrack data set establishment method in the S1 comprises the following specific steps: s1.1, collecting jpg-format original image data of cracks and peeling damage on the surface of a concrete structure; s1.2, in Photoshop software, manually marking the cracks and the peeling damage pixels in the original image collected in the S1.1, and manufacturing a png-format label image of the original image of the cracks and the peeling damage collected in the S1.1; in the label image, the background pixel, the crack pixel and the peeling pixel are respectively represented by 0, 1 and 2; s1.3, adjusting the sizes of the crack and the peeling original image collected in the S1.1 and the label image corresponding to the crack and the peeling original image to be 504 multiplied by 376 pixel size; s1.4, randomly selecting 80% of original images and label images corresponding to the original images from the crack and peeling original images and the label images corresponding to the original images after the size adjustment in S1.3 to serve as a training set of the deep learning semantic segmentation network model, and the remaining 80% of original images and the label images corresponding to the original images to serve as a verification set; the specific steps of constructing the deep learning semantic segmentation full convolution network for identifying the concrete structure surface cracks and the peeling damage pixels in the S2 are as follows: s2.1, establishing a deep learning semantic segmentation full convolution network for recognizing concrete structure surface cracks and peeling damage pixels based on a DenseNet-121 convolution neural network; s2.2, initializing weight and bias parameters of each layer in the established deep learning semantic segmentation full convolution network; s2.3, setting learning rate, momentum and weight attenuation super-parameters during training and verification of the established deep learning semantic segmentation full-convolution network; the specific steps of training and verifying the deep learning semantic segmentation network constructed by the S2 by adopting the data set constructed by the S1 in the S3 are as follows: s3.1, training a deep learning semantic segmentation full convolution network constructed in the S2 by adopting the training set constructed in the S1.4; s3.2, inserting a verification process in the training process of the deep learning semantic segmentation full convolution network in the S3.1, and verifying the deep learning semantic segmentation full convolution network model obtained in the training process by adopting a verification set established in the S1.4; the specific steps of adopting the deep learning semantic segmentation network trained and verified by the step S3 to identify the cracks and the peeling damage pixels in the surface image of the bent reinforced concrete simply supported beam in the step S4 are as follows: s4.1, dividing the bent reinforced concrete simply supported beam into a midspan and a side span by taking a left support and a right support as end points and two concentrated symmetrical load acting points as intermediate points, and marking the area range of the midspan and the side span on the surface of the bent reinforced concrete simply supported beam; s4.2, after the bent reinforced concrete simply supported beam is damaged, an intelligent mobile phone is adopted to collect images of a span area where damage exists on the surface of the damaged bent reinforced concrete simply supported beam, and the surface area of the bent reinforced concrete simply supported beam contained in the images is recorded, wherein the surface area is as follows: midspan or side span; s4.3, cutting out a midspan or side span region part in the damaged bent reinforced concrete simply supported beam surface image acquired in the S4.2 according to the mark made in the S4.1, and adjusting the length and width pixel sizes of the cut midspan or side span region image to be integral multiples of 504 multiplied by 376 pixels; s4.4, scanning the middle-span or side-span region image of the damaged bent reinforced concrete simply supported beam surface subjected to size adjustment in the step S4.3 by adopting a rectangular sliding window with the size of 504 multiplied by 376 pixels sequentially from left to right and from top to bottom; in the scanning process, when the window slides to a certain position, a trained deep learning semantic segmentation network model in the S3 is applied to identify cracks and flaking pixels in a small image at the position; in the image in the identification result, the background, the crack and the peeling pixel are respectively represented by black, white and green; and S4.5, adjusting the size of the image of the identification result of S4.4 to the original size of the partial image of the midspan or side span region cut out in S4.3.
2. The deep learning-based method for identifying the failure mode of the bent reinforced concrete simply supported beam, which is characterized by comprising the following steps of: the specific steps for extracting the damage characteristics of the positive section or the inclined section of the bent reinforced concrete simply supported beam in the image identified in the step S5 are as follows: s5.1, analyzing the identification result image after the size adjustment of S4.5 according to the range of the surface area of the bent reinforced concrete simply supported beam contained in the acquired image recorded in S4.2, wherein the range is a midspan or a side span, and if the acquired image of S4.2 is a midspan area image of the bent reinforced concrete simply supported beam, judging whether a main crack exists in a tension area of the beam and whether peeling damage exists in a compression area of the beam; s5.2, according to the range of the surface area of the bent reinforced concrete simply supported beam contained in the acquired image recorded in S4.2, wherein the range is a midspan or a side span, if the image acquired in S4.2 is a side span area image of the bent reinforced concrete simply supported beam, analyzing the identification result image after the size adjustment in S4.5, performing straight line fitting on pixel coordinates of a crack in the identification result image, and calculating an absolute value of a slope of the fitted straight line, wherein the absolute value is a shear span ratio of the bent reinforced concrete simply supported beam calculated according to the crack.
3. The deep learning-based method for identifying the failure mode of the bent reinforced concrete simply supported beam, which is characterized by comprising the following steps of: the specific steps of the S6 for identifying the damage mode of the bent reinforced concrete simply supported beam according to the damage characteristics of the normal section or the inclined section extracted in the S5 are as follows: s6.1, if the image acquired in the S4.2 is a midspan area image of the simply supported beam of the bent reinforced concrete, identifying a positive section damage mode of the simply supported beam of the bent reinforced concrete according to the judging result of whether a main crack exists in a tension area of the beam and whether peeling damage exists in a compression area of the beam in the S5.1; and S6.2, if the image acquired in the S4.2 is an edge span area image of the bent reinforced concrete simply supported beam, identifying the inclined section damage mode of the bent reinforced concrete simply supported beam according to the range of the shear span ratio of the bent reinforced concrete simply supported beam calculated according to the crack in the S5.2.

Description

Method for identifying failure mode of flexural reinforced concrete simply supported beam based on deep learning Technical Field The invention relates to the field of failure mode identification of reinforced concrete members, in particular to a failure mode identification method of a bent reinforced concrete simply supported beam based on deep learning. Background The reinforced concrete structure is widely applied to various large-scale engineering structures. For the Gramineae of civil engineering specialty, a specialized course is very important in the concrete structural design principle. In the teaching process of the course, a flexural reinforced concrete simply supported beam damage experiment is carried out, and damage mode identification is carried out according to the damaged surface characteristics of the beam, so that the method is a key teaching link of the course. Through the damage experiment of the bent reinforced concrete simply supported beam, students in civil engineering profession can recognize the relation between the factors such as the reinforcement ratio, the concrete strength grade, the section form and the like of the bent reinforced concrete simply supported beam and the damage mode. At present, when a flexural reinforced concrete simply supported beam damage experiment is carried out, a manual method is mainly adopted for identifying a damage mode. When the bent reinforced concrete simply supported beam generates an initial crack, the initial crack needs to be observed in a visual manner. After the positive section of the bent reinforced concrete simply supported beam is damaged, whether a main crack exists in a midspan tension zone of the beam and whether concrete in a compression zone is crushed to form flaking or not needs to be observed manually, and the positive section proper-rib, the few-rib and the super-rib damage modes of the bent reinforced concrete simply supported beam are identified according to different combinations of the two phenomena. After the normal section of the bent reinforced concrete simply supported beam is damaged, the form of the beam oblique crack needs to be analyzed, and the oblique section shearing and pressing, oblique pressing and oblique pulling damage modes of the bent reinforced concrete simply supported beam are identified according to analysis results. Although the manual method can effectively identify the damage mode of the bent reinforced concrete simply supported beam, the method requires personnel with professional knowledge and expensive equipment and is time-consuming and labor-consuming. Disclosure of Invention The invention aims to provide a method for identifying a failure mode of a bent reinforced concrete simply supported beam based on deep learning, which realizes high-precision and automatic identification of the failure mode of the bent reinforced concrete simply supported beam. The purpose of the invention is realized in the following way: the method for identifying the damage mode of the bent reinforced concrete simply supported beam based on deep learning comprises the following steps: s1, establishing a reinforced concrete structure surface crack and peeling damage semantic segmentation data set; s2, constructing a deep learning semantic segmentation full convolution network for identifying surface cracks and peeling damage pixels of the reinforced concrete structure; s3, training and verifying the deep learning semantic segmentation full convolution network constructed in the S2 by adopting the data set constructed in the S1; s4, performing crack and peeling damage pixel identification in the surface image of the bent reinforced concrete simply supported beam by adopting the deep learning semantic segmentation full convolution network model trained and verified in the S3; s5, extracting the damaged characteristics of the positive section or the inclined section of the bent reinforced concrete simply supported beam in the image identified in the S4; s6, identifying the damage mode of the bent reinforced concrete simply supported beam according to the damage characteristics of the normal section or the inclined section extracted in the S5. Further, the specific steps of establishing the microcrack data set on the surface of the concrete structure in the step S1 are as follows: s1.1, collecting jpg-format original image data of cracks and peeling damage on the surface of a concrete structure; s1.2, in Photoshop software, manually marking the cracks and the peeling damage pixels in the original image collected in the S1.1, and manufacturing a png-format label image of the original image of the cracks and the peeling damage collected in the S1.1; in the label image, the background pixel, the crack pixel and the peeling pixel are respectively represented by 0, 1 and 2; s1.3, adjusting the sizes of the crack and the peeling original image collected in the S1.1 and the label image corresponding to the crack and the peeling ori