Anomaly Detection in Time Series using Generative Adversarial Networks

Abstract

Generative Adversarial Networks (GANs) have been applied to an increasing amount of tasks, especially related to image data. A comparably recent advance was their application to the domain of anomaly detection in images and, even more recently, on spatiotemporal data. In this work, a recurrent GAN (RGAN) is applied on cardiovascular data from the MIT-BIH dataset to learn the natural variety of normal sinus rhythms in a healthy individual. The generator is used to reconstruct samples using differently parameterized levels of similarity and thresholds. We find that solely using the generator already allows a surprisingly good anomaly detection performance. Furthermore, we discuss adding the discriminator, which might significantly improve the performance. Future work also includes only using the discriminator, minimizing the time required for inference, which is important for streaming data.