基于预训练交互式图神经网络的多元时间序列异常检测  

Multivariate time series anomaly detection using pre-training based interactive graph neural network

作　　者：王春枝[1] 邢绍文 高榕 严灵毓[1] WANG Chunzhi;XING Shaowen;GAO Rong;YAN Lingyu(School of Computer Science,Hubei University of Technology,Wuhan 430072,China)

机构地区：[1]湖北工业大学计算机学院,武汉430072

出　　处：《中南民族大学学报（自然科学版）》2023年第4期541-550,共10页Journal of South-Central University for Nationalities：Natural Science Edition

基　　金：湖北省重点研发资助项目(2020BAB012)。

摘　　要：多元时间序列异常检测是数据挖掘领域中的一项重要应用.基于深度学习的异常检测方法已经取得了重大进展,但其仍然存在一定的局限性.首先,是它们假设训练数据仅由正常数据组成,而忽略了异常数据可能导致的不可预测性;其次,大部分方法并未考虑到时间序列的独特特性.为了解决上述问题,基于预训练提出了一种新颖的用于多元时间序列的异常检测框架.框架由预训练模块和预测模块组成,首先预训练模块通过学习时间序列的密集向量表示,增强其可预测性,然后预测模块中充分利用时间序列的独特特性捕获其时间依赖.通过广泛的实验证明了所提出的模型的有效性,在三个真实数据集上均显著优于最先进的模型.Multivariate time series anomaly detection is an essential application in data mining.Recent deep learning-based anomaly detection methods have made significant progress,but they still have some limitations.Firstly,they assume that the training data consists of only normal data and ignore the unpredictability that may result from anomalous data.Secondly,most of the methods do not take into account the unique characteristics of time series.To address the above problems,a novel anomaly detection framework for multivariate time series is proposed based on pre-training.The framework consists of a pre-training module and a prediction module.Firstly,the pre-training module enhances the predictability of the time series by learning its dense vector representation.Then the unique properties of the time series are fully exploited in the prediction module to capture its time dependence.The effectiveness of the proposed model is demonstrated through extensive experiments,significantly outperforming the state-of-the-art model on all three real data sets.

关 键 词：时间序列 异常检测 预训练 Transformer编码器 图神经网络 

分 类 号：TP183[自动化与计算机技术—控制理论与控制工程]