LSTM-GAN:融合GAN和Bi-LSTM 的无监督时间序列异常检测

LSTM-GAN:Unsupervised Anomaly Detection for Time Series Fusion of GAN and Bi-LSTM

多元时间序列数据的异常检测旨在发现对应时序特征中不符合一般规律的特异性模式,识别特定时间步长中的异常状态.针对多元时序数据时间依赖性建模难以及数据维度不断增加导致难以有效进行异常检测等问题,本文以自编码器为基础,融合生成对抗网络(Generative Adversarial Network,GAN)和双向长短期记忆神经网络(Bi-directional Long Short-Term Memory,Bi-LSTM),提出了一种无监督异常检测模型LSTM-GAN,该模型在每一轮训练中,以迭代的方式重构正常数据,通过GAN来放大异常,Bi-LSTM来捕获时间特性,训练完成后的模型用于时序数据的异常检测.本文在4个公开数据集上和几种先进同类方法进行了对比实验,实验结果表明,LSTM-GAN的检测性能提升了4.4%~16.6%,在IT数据集SMD中的模型检测F1分数达到0.9672,实现了高效的时序数据异常检测.

Anomaly detection of multivariate time series data aims to discover specific patterns in the corresponding time series features that do not conform to the general pattern and identify anomalous states in a specific time step.To address the problems of time dependence modeling of data and the difficulty of effective detection due to the increasing dimensionality of data,this paper proposes an unsupervised anomaly detection model LSTM-GAN based on Autoencoder fusing Generative Adversarial Network(GAN)and Bi-directional Long Short-Term Memory(Bi-LSTM),which reconstructs normal data in an iterative manner,amplifies anomalies by GAN,and captures temporal characteristics by Bi-LSTM,the completed model is used for anomaly detection.We conducted comparison experiments on four publicly available datasets,and the experimental results show that the performance of LSTM-GAN is improved by 4.4%~16.6%,and the F1 score of the model in SMD reaches 0.9672,which achieves efficient anomaly detection of temporal data.