DeepAnom: A Hybrid Time GAN and LSTM Autoencoder Framework for Time-Series Anomaly Detection in IIoT Systems

Abstract

Data represented in the form of a time series appears in many fields, including finances or health care, or in industries where the role of detecting anomalies becomes crucial for guaranteeing stability. However, it appears that the reason behind such high importance of anomaly detection comes from the fact that its lack might result in many problems. Unfortunately, anomaly detection models often fail due to the problem of lack of correctly labeled datasets for training purposes. For this reason, it is recommended to use Time GAN model, i.e., time series generative adversarial network for creating synthetic data sets which would help to improve anomaly detection. The idea of Time GAN lies in combining the methods of adversarial training and supervised learning, thus adding to the model both the temporal component and relevant features. Such a method allows for learning the time series patterns without just copying them. When a dataset is created using a Time GAN model, it can be included into the existing datasets to make further improvements to training. Experiments have shown that incorporating synthetic data into a model improves anomaly detection accuracy and robustness.