A Novel LSTM-GAN Algorithm for Time Series Anomaly Detection

Abstract

Time series anomaly detection is an important part of Prognostic and Health Management (PHM), and has been widely studied and followed with interest. The data with time series features often has non-stationary properties, and its fluctuation amplitude changes with time. Traditional anomaly detection algorithms can achieve the detection of shallow level anomalies when facing such data, however they fail to detect outliers on deep features of time series data. The gate structure of the long short-term memory network (LSTM) shows obvious advantages in processing time series data, while the confrontation training of generative adversarial network (GAN) performs well in detecting and acquiring deep features of data. Therefore, this paper focuses on the anomaly detection of time series data with the fusion model of LSTM and GANs, which is named the LSTM-GAN, and the performance of the algorithm is verified from two sets of time series data. The experimental results demonstrated that the proposed algorithm achieved superior performance in processing time series data compared to conventional algorithms. The research content of this article has profound guiding significance for time series anomaly detection.