Surface Defect Recognition And System Development Of Hot-rolled Strip Based On Convolutional Neural Network

Hot-rolled strip is one of the most important products in the steel industry,and it has a very wide range of applications in the field of manufacturing.In the production process of the hot-rolled strip,due to performance degradation of rolling equipment and many other factors,leading to the different types of surface defects,such as inclusion,warping,and scratch,which bring the economic toss to the steel enterprises,and to a great extent,they restrict the development of the steel enterprises.Meanwhile,the traditional manual testing methods of the steel enterprises can’t guarantee the requirements of fast rhythm production,and the detection results of using the surface detection device are not very ideal.This thesis analyses the deep learning which has developed rapidly in the field of the artificial intelligence research in recent years,and applies the convolutional neural network in the field of the deep learning to identifying the surface defects of the hot-rolled strip.The main research work is as follows:1)This thesis summarizes the domestic and abroad research status of the surface defect detection technology of the hot-rolled strip,deep learning,and analyzes some problems faced by the traditional manual test and equipment detection methods.2)This thesis studied the features of the convolutional neural network compared with the traditional artificial neural network,and the conduction calculation process of the forward propagation and back propagation deeply,and summarizes the relative issues needing to be concerned in the process of training the convolutional neural network,and some effective methods preventing the training from overfitting.3)In allusion to ten hot-rolled strip surface defects,this thesis sets up an image sanple database.And in allusion to the database,this thesis establishes an improved model based on the classical convolutional neural network model,and it makes good classification results under using the accelerated training of GPU and CUDNN.4)This thesis extracts the well-trained top feature of the convolutional neural network model as the feature vector to retrieve the images of the hot-rolled strip surface defects.This method does not require people to design the features to describe the operators,and it learns the features of the images directly,which is more robust and of high retrieval accuracy.5)This thesis designs and develops the classification and retrieval system of the hot-rolled strip based on the convolutional neural network with the background of the actual demand and the core of the convolutional neural network.The system uses GUI MATLAB to make the front desk user interface,establishes the convolutional neural network based on the deep learning library MATCONVNET,and applies the acceleration framework GPU and CUDNN for network training.The system can provide users with three key functions:model training,prediction classification and image retrieval.Through the application of practical problems,the model and the system are verified,and the results show that the system is easy to operate,which is in line with the actual production requirements and can provide users with effective help.