Identification Of Corroded Cracks In Reinforced Concrete Based On Deep Learning Convolutional Neural Network

As the most widely used building material,the durability of reinforced concrete structure will affect the appearance and bearing capacity of the building.As a kind of durability problem,the corrosion of reinforced concrete structure will cause the volume expansion of reinforcement,lead to the development of concrete cracks and eventually lead to the fall of concrete protective layer,resulting in structural damage and decline of bearing capacity.Therefore,it is of great significance to timely detect concrete cracks caused by steel corrosion.Traditional monitoring needs to be completed manually at regular intervals,but there are some problems such as poor safety,high cost and low efficiency.With the development of computer,some damage detecti on methods based on computer vision are mainly image processing technology,which can detect the surface defects of concrete.However,the current image processing technology also has some limitations,such as light and noise have a great impact on the results.Deep learning is a learning algorithm to extract features by building a complex level neural network and using a large amount of data to train and learn the neural network.Convolutional Neural Network(CNN)combines image processing technology with deep learning and trains the Network model by using a large scale data set,which can save time,improve detection efficiency and recognition accuracy.In this paper,CNN is applied to the identification and classification of reinforced concrete corrosion cracks,a data set of concrete cracks caused by reinforcement corrosion is constructed,and a model of reinforced concrete corrosion crack identification based on deep learning CNN is proposed.Compared with two traditional detection methods,the accuracy and testability of the model are verified.Based on the theories related to deep learning,this paper studies the basic principles of convolutional layer,pooling layer,full connection layer and Softmax layer according to the basic structure of convolutio nal neural network.According to the working principle of convolutional layer and pooling layer,the frame structure of classic CNN is summarized,and the design rule of CNN frame for image classification is obtained.In order to obtain high accuracy of CN N,a large and real data set was constructed by means of network search,self-shooting and electrification to accelerate corrosion test of reinforcement.The data set includes the pictures of concrete cracking caused by steel corrosion,concrete cracks caused by other reasons and complete concrete pictures.The difference between concrete cracks caused by steel corrosion and concrete cracks caused by other reasons is mainly reflected in the cracks along the reinforcement,accompanied by the precipitation of yellow-brown corrosion products.A standard for image acquisition was set up.After the acquisition,image data was enhanced by Open CV methods such as clipping,rotation and color transformation.It is divided into training set and test set according to corresponding scale.This paper uses Tensor Flow learning framework to construct neural network classification model,which is divided into four parts: Data set is transformed into Tensor Flow special format TFRecords as neural network input;The CNN frame SCNet is designed to classify the concrete crack caused by the corrosion of steel bar and the concrete crack caused by other reasons The training and testing program is constructed by Tensor Flow with the parameters of activation function,loss function,learning rate and back propagation.By training the training sample set,the model is tested on the training sample set and the test sample set,and the training precision is 98.5% and the test precision is 96.8%,therefore,it can meet the requirements for the identification of corrosion cracks in reinforced concrete.According to the training accuracy and test accuracy of the model,the network structure and network parameters are adjusted,and the influence of different neural network structures,different o ptimization methods and different super parameters on the classification accuracy is discussed.In order to study the changes in light intensity,light and shade and image distortion and so on all sorts of problems and the noise influence this method measu rability,will build the SCNet model and two kinds of traditional edge detection methods Sobel edge detection method and Canny edge detection method,analysis of test results of three methods under different environment,comparative and discussed in this paper,build CNN under different environmental conditions of reinforced concrete corrosion fracture identification of measurability and superiority.