基于LSTM-WGAN的时间序列数据异常检测  

Anomaly Detection for Time Series Data with LSTM-WGAN

作　　者：郑圣彬 谢加良[1] 张东晓[1] ZHENG Shengbin;XIE Jialiang;ZHANG Dongxiao(School of Science,Jimei University,Xiamen 361021,China)

机构地区：[1]集美大学理学院,福建厦门361021

出　　处：《福建师范大学学报（自然科学版）》2024年第2期36-45,共10页Journal of Fujian Normal University：Natural Science Edition

基　　金：国家自然科学基金资助项目(12271211,12071179);福建省自然基金资助项目(2020J01710,2021J01861)。

摘　　要：在时序数据中发现隐藏的异常行为或事件,可以保障生产安全,具有重要意义。目前的异常检测模型存在训练不稳定、容易产生梯度消失的问题,影响异常检测效果,针对该问题,提出一种LSTM-WGAN模型,WGAN负责捕获变量之间的潜在关联,进一步提升了LSTM的检测能力。同时,以Wasserstein距离代替交叉熵损失训练判别器和生成器,结合重构损失以及判别损失实现异常检测。在NAB公开数据集上的实验结果表明LSTM-WGAN相较于基准模型在准确率、召回率以及F1得分上都有较大幅度的提升。Detecting hidden anomalies or events in temporal data is crucial for ensuring production safety.However,the current anomaly detection models are prone to unstable training and gradient vanishing,which impair the effectiveness of anomaly detection.To address this issue,this paper proposes an LSTM-WGAN model,where WGAN captures the potential correlations between variables to enhance LSTM s detection ability.Moreover,the LSTM-WGAN model uses Wasserstein distance instead of cross-entropy loss for training both the discriminator and generator,integrating reconstruction loss and discrimination loss for effective anomaly detection.The experimental results on the NAB public dataset demonstrate that LSTM-WGAN outperforms the benchmark model in terms of accuracy,recall,and F1-score.

关 键 词：异常检测 时间序列 生成对抗网络 长短期记忆网络 Wasserstein距离 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]