Research And Application Of Generative Adversarial Network’s Generative Stability

As a remarkable generative model,generative adversarial networks are increasingly becoming one of the hot research area of deep learning research since it was proposed.Generative adversarial networks can be used in image super-resolution,data enhancement,target detection and image restoration.Although it is widely applied in other fields,the effect achieved is not as good as that of the image.Generative adversarial networks currently have training instability.When the training of the generator and the generator does not reach the ideal synchronization,the discriminator will lose its function.GAN may also have the phenomenon of mode collapse.If the GAN’s loss function and other parameter settings are not reasonable,then the gradient will disappear and the gradient will explode.Up to now,the unstable training and model collapse of GAN have not been solved well.This article will focus on the network structure and loss function of the generative countermeasure network,analyze and improve the stability of the generative countermeasure network,and use it to other areas.The main contribution of the article is:1.Make the generator structure of the GAN more effective.In the generator,a convolutional network is applied to obtain characteristics of the data,and a LSTM is applied at the same time to get the characteristics of the time series,so that the generator can generate better data samples.2.Improve the loss of the generated adversarial network.Through the study of the loss in GAN,the origins of the crumbling of network and what makes generator unstable.A new loss function of the generative adversarial network is proposed,Minkowski distance is applied in objective function,so that the training of the generative adversarial network can be well behaved.3.Apply the improved generative confrontation network to lithium battery life prediction.Through the proposed new loss function,the loss function is added to the GAN’s generator,and applied to the prediction of lithium battery life.The performance in this work show that on different data sets,the performance in this work in this paper can be better in estimating the life of lithium batteries,the method used in this work tends to have lower MAE,MAPE,RMSE in the estimation of lithium battery life.