Data Augmentation Method Based On Generative Adversarial Network And Its Application

In recent years,with the rapid development of China’s manufacturing industry,the country needs to continuously improve its ability to design and develop complex products.Due to various factors such as technological iteration,changes in customer demand,and changes in component supply chains,design changes for complex products are pervasive throughout the entire development cycle.Accurately evaluating the impact of design changes on complex product design is of great significance to improving the development level of the manufacturing industry.With the advancement of deep learning technology,data-driven learning algorithms have been widely applied in various fields of society.However,there is a problem of insufficient historical design change data in the engineering application of product design change analysis.Therefore,this article focuses on the problem of insufficient data in complex product design changes and researches the application of the data augmentation method based on generative adversarial network in predicting the intensity of design changes for televisions.The main research work of this thesis is as follows:(1)To address the issue of insufficient training data for neural network,a latent feature fusion generative adversarial network(LFFGAN)is designed in this thesis.To improve the ability of the generative adversarial network to generate high-quality and diverse data in tasks with limited data,this thesis conducts in-depth research on latent feature interpolation methods.Multiple latent feature values of real data are extracted using an encoder,and a latent feature similarity fusion method is proposed to fuse the features.The fused latent features are then used as inputs to the generator instead of random noise,effectively reducing the difficulty of traditional generative adversarial network in learning the real sample distribution using random noise,thereby improving the learning ability of the generator.To ensure the quality of generated data,a reconstruction loss function is designed using the similarity relationship between latent features to further enhance data augmentation effects.The LFFGAN proposed in this thesis is proven to effectively improve downstream task model performance in tasks with limited data through data augmentation in image classification tasks.(2)Based on the research of(1),propose a complex product design change intensity prediction method based on data augmentation to analyze issues regarding product design change.To address the problem of data scarcity in design change analysis,the thesis uses one-hot encoding to process the qualitative data in product design change data,which facilitates the use of an improved generative adversarial network(GAN)for data augmentation and deep learning algorithms for predicting design change intensity based on part parameters.By modeling complex products using historical change data,the thesis accurately identifies the core components of complex products based on the global,local,and positional attributes of components,using the Technique for Order of Preference by Similarity to an Ideal Solution(TOPSIS)and entropy methods.Based on the characteristics of design change data,the thesis proposes a K-nearest neighbor and latent feature fusion GAN(K-LFFGAN)to augment components design change data.To improve the accuracy of design change intensity prediction,the thesis uses ensemble learning and convolutional neural network to build an ensemble convolutional neural network.Finally,the effectiveness of the proposed method was validated in the analysis of a design change case for a certain model of Skyworth television.