Time Series Data Augmentation Algorithm Combining Deep Metric Learning and Variational Encoder

Aiming at the challenges of current time series data generation and the slow speed of the generation framework based on generative adversarial networks, we propose a novel architecture for comprehensively generating time series data using variational autoencoders. Specifically: study how to eliminate the heteroskedasticity and non-stationarity of the data, so that when the data is applied to the encoder, it can better conform to the normal distribution assumption; then introduce deep metric learning, based on metric learning to build a more suitable data enhancement discriminative VAE latent space; based on the above two studies, they are integrated to obtain a novel variational autoencoder framework. We verified the necessity of each module through ablation studies; the quality of data generation was evaluated by the similarity and predictability of the three multivariate datasets. Our similarity test results show that the VAE method can accurately represent the temporal properties of the original data. In the prediction task using the generated data, the proposed VAE system can obtain the best results compared with other SOTA algorithms, and the average speed is increased by about 20.2%.