Research On Prediction Method Of Grain Boundary Corrosion Resistance Based On Deep Learning

The traditional prediction of corrosion resistance of white copper alloy is based on manual selection of relatively simple and significant features or physical experiment method.However,these methods have some problems such as strong subjective factors,low prediction accuracy and high cost.How to extract and predict the features of grain boundary images quickly and accurately has become an urgent problem.Deep learning has become a hot topic in the field of computer application.And it has achieved good results in major industrial fields.In recent years,the application range of copper alloys has become wider and wider,and more attention has been paid to its service life and its corrosion resistance to seawater when used as marine materials.There is a specific relationship between the corrosion resistance of the alloy and grain boundary microstructure.In order to better improve the quality of white copper alloy and study the structural characteristics of grain boundary,Further analysis of grain boundary images is needed.In view of the wide application of deep learning in images,a prediction model of grain boundary corrosion resistance based on optimized convolutional neural network was proposed.Furthermore,a prediction model of grain boundary corrosion resistance based on improved residual network was proposed according to deep network.The main work contents are as follows.(1)A prediction model of grain boundary corrosion resistance based on optimized convolutional neural network was proposed.In this paper,the convolutional neural network in deep learning is introduced into the corrosion prediction of grain boundary images.The convolutional neural network structure was constructed according to the characteristics of grain boundary images.The convolutional neural network is further optimized by using the characteristics of its structure.The traditional convolution operation has the problems of high parameter consumption and long learning time.To solve these problems,a deep separable convolution is proposed to replace the traditional convolution process.Theoretically and experimentally,it is proved that this substitution can effectively reduce the consumption of parameters and improve the accuracy of the model.By selecting the operation strategy of the pooling layer,the learning effect of the model is further improved.It can reduce the loss of important feature information of grain boundary,and more effective feature information can be extracted to improve the accuracy of model prediction.In addition,a multi-layer feature fusion learning module is proposed to fuse and then learn the grain boundary features learned at different layers.Ensure the richness of feature information extracted from the model.Through experiment comparison,it is verified that the proposed improvement and optimization have better effect on the network model.(2)A prediction model of grain boundary corrosion resistance based on improved residual network was proposed.The features of grain boundary images were extracted effectively by establishing three different learning channels.Then the characteristic information learned by the three channels is unified and integrated.Then the integrated learning module is used to classify the grain boundary images.It is found that the residual network is the most suitable for grain boundary image training by building a variety of convolutional network structures.Through experimental comparison,it is verified that the proposed model has a better prediction effect on grain boundary image classification.Finally,an improved residual network was proposed to strongly extract grain boundary features through multiple channels,and ensemble learning was used to achieve higher classification prediction accuracy.