Abstract
This paper describes a system for detecting anomalies such as outliers and periodic fluctuations in time-series data, including data acquired by satellites, and how to improve the accuracy of the system. Although there are various methods for anomaly detection, machine learning methods have been actively studied in recent years, unsupervised learning methods that do not provide correct labels are rare due to their low accuracy. We developed an anomaly detection system based on unsupervised learning using TadGAN, an adversarial generative network, and solved its drawbacks by performing hyperparameter tuning using FlexHyperband. This system enables fast anomaly detection on large data sets without the need for human intervention.
