Synthesizing Vibration Signals Using Generative Adversarial Networks

This work evaluates the applicability of GANs in generating synthetic time series data. Hypoid gear vibration data was used as experimental data to test models. GANs for time series can be divided into two different types: discrete type (discrete time series) and continuous type (continuous time series). In time series, there are complex time-dependent features and their attributes, for example, when using biometric data, ECG characteristics will depend on the age and health of the individual. TimeGAN is a generative time series model that combines unsupervised learning using GAN with supervised autoregressive learning. TimeGAN model performace was evaluated using predictive score and visualization using PCA and tSNE analysis for original and synthetic data. The best predictive score was 0.062, however visual evaluation using tSNE and PCA analysis revealed that generated data was not sufficiently similar to the real. It can be concluded that the application of GANs to time series presents difficulties that are not present in image generation models. Continuing this work could lead to testing other time series GANs in the future, e.g. DoppelGANger whose synthetic data structure does not require complex reconstruction.