基于生成式对抗网络的自监督多元时间序列异常检测方法  

Self-supervised multivariate time series anomaly detection based on GAN

作　　者：周业瀚 沈子钰 周清 李云[1,2] Yehan Zhou;Ziyu Shen;Qing Zhou;Yun Li(College of Computer Science,Nanjing University of Posts and Telecommunications,Nanjing,210023,China;Jiangsu Key Laboratory of Big Data Security&Intelligent Processing,Nanjing,210023,China)

机构地区：[1]南京邮电大学计算机学院,南京210023 [2]江苏省大数据安全与智能处理重点实验室,南京210023

出　　处：《南京大学学报（自然科学版）》2023年第2期256-262,共7页Journal of Nanjing University（Natural Science）

基　　金：江苏省研究生科研创新计划(KYCX_0760)。

摘　　要：异常检测是数据挖掘的重要研究方向之一.工业设备的各项指标以多元时间序列的形式被传感器监测,多元时间序列的异常检测对保障安全和提高服务质量至关重要,但是异常的定义相对模糊,具有异常标签的数据很稀少.此外,多元时间序列具有复杂的时间依赖性和随机性,使异常检测存在许多问题.提出CPCGAN模型,使用自监督学习的方法对多元时序数据进行异常检测.首先使用对比学习的方法得到多元时序数据的表示向量,再将具有先验信息的表示向量作为输入用来训练生成式对抗网络,通过生成式对抗网络的重构误差来确定异常.在五个数据集上与五种无监督异常检测方法进行对比,实验结果证明提出的方法能有效地检测两类异常,并且,在大多数数据集上的表现更好.Anomaly detection is one of the important research directions of data mining.The indicators of industrial devices are monitored by sensors in the form of multivariate time series.Anomaly detection of multivariate time series is critical for security and improving service quality.However,the definition of anomalies is relatively vague and the data with anomalous labels is rare.Also,multivariate time series have complex time dependence and stochasticity which makes anomaly detection many issues to be settled.In this paper,we propose CPCGAN,a self-supervised learning method,to perform anomaly detection on multivariate time series data.Our main idea is to obtain the representation vector of multivariate time series data by using the contrastive learning method.We use the representation vector with prior information as input when training the generative adversarial network.The reconstruction error of the generative adversarial network is used to determine anomalies.We compare our method with five unsupervised anomaly detection methods on five datasets.Experimental results show that our method is effective at detecting both types of anomalies and performs better on most datasets compared with other methods.

关 键 词：异常检测 多元时间序列 自监督学习 对比学习 生成式对抗网络 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]