Synthesis Algorithm of Data in Synchronization Network Based on Modified Variational Autoencoder

In this paper aiming to generate plentiful testing samples on demand in synchronization network, a synthesis method with machine learning is proposed inspired by generative model and loss function of style transfer. Firstly equivalent relationship between synthesis data and phase response is analyzed, based on which the modified VAE (Variational Auto Encoder) framework is designed where attribute features of phase response for synthesis data are represented in latent space, grafting a style evaluation network for the purpose that the diversity of style for data generation is quantified and evaluated. Secondly synthesis algorithm is implemented with a novel scheme to operate and select attribute feature in latent space under proposed criteria as posterior processing to balance the style diversity and fitness with the category of anchor sample from standard template. Moreover the experimental results verify that the proposed algorithm show some advantages in the aspect of generation speed, richness of samples and accuracy of fitness for objective on demand, comparing to conventional method such as simulated annealing or genetic algorithm. Finally two potential application scenarios adopting this synthesis algorithm are recommended in synchronization network.