Application Of Convolutional Neural Network In Corrosion Resistance Prediction Of Alloys

Artificial neural network is a model commonly used in the field of computer application.It is inspired by the structure and function of biological neural network and brain,convolutional neural network is one of them,because of its adaptability and generalization performance,it has become a research hotspot in the field of machine learning.With the rapid advancement of the country's "Ocean Power" strategy,the ship's main material B10 copper-nickel alloy is more and more concerned with the problem of its service life due to the tendency to corrode in the process of serving in the seawater environment.The properties of the material are closely related to its microstructure.For better design and more precise control of the grain boundary structure,it is indispensable to carry out a more detailed and reliable analysis of the grain boundary microstructure.In this paper,a grain boundary connectivity model based on image analysis is proposed for the defects of B10 copper-nickel alloy microstructure analysis model and the single feature.And guided by this model,A prediction model for corrosion resistance of B10 copper-nickel alloy based on optimized convolutional neural network is proposed.At the same time,three improved strategies are proposed for the large amount of convolution calculation in traditional neural networks,the loss of pooling operation features and the lack of high-level features.The main tasks are:(1)A prediction model for corrosion resistance of B10 copper-nickel alloy based on image analysis is proposed.Based on the analysis and calculation of grain boundary information in grain boundary images,a quantitative model of grain boundary connectivity is established by using two key attributes of connectivity frequency and angle between grains.In order to obtain the detailed distribution information of the two attributes,three algorithms are presented in turn: grain boundary refinement,intersection extraction and classification,and boundary angle calculation.The weights of the two attributes in the connectivity model are established by experiments.In order to verify the predictive effect of the model,the corrosion resistance prediction of the established grain boundary connectivity model is verified with the results of physical experiments,and the minimum size of the grain boundary image which can be used for convolution neural network training is calculated according to the model.(2)A prediction model of corrosion resistance of B10 copper-nickel alloy based on optimal convolution neural network is proposed.A convolution neural network architecture for grain boundary image characteristics is presented.A step-by-step convolution operation is proposed based on the analysis of the traditional convolution operation process.It is proved theoretically and experimentally that the operation can reduce the consumption of additional parameters.The loss of feature information is reduced by the proposed learning single channel pooling operation,and the multi-layer feature fusion learning strategy is proposed to diversify the feature expression and improve the prediction accuracy of the model.In order to verify the applicability and accuracy of the model,three control variable experiments were conducted to test and compare the improved operation effect.The accuracy of image classification effect and prediction ability of grain boundary corrosion resistance of the improved model on open data sets were verified by two other experiments compared with traditional convolution neural network model.