摘要
为了建模多维时间序列(multivariate time series,MTS)中复杂的时序依赖性与随机性,并实现对它的无监督异常检测这一工业机器或互联网基础设施设备质量管理中的关键任务,本文提出了一种堆叠式对抗变分循环神经网络(stacked adversarial variational recurrent neural network,SaVRNN).SaVRNN是一个层次化概率动态模型,它将层次化概率生成模型、堆叠式循环结构和多层对抗优化方式整合到一个联合贝叶斯框架下.具体来说,SaVRNN核心思想是利用堆叠循环结构捕捉多层次与长距离的时序相关性,利用层次化的概率生成操作建模多层的结构特性,进而实现对多维时间序列正常模式的学习,最后通过重构的概率来判断异常模式.为了实现模型的高效推理,本文创新性地提出了一种向上–向下对抗变分推理方案,实现对隐层变量后验的准确估计.针对多层对抗网络中难以实现生成器与判别器的更新平衡导致的SaVRNN训练困难的问题,本文基于条件传输(conditional transport,CT)提出了一个新的优化方法.最后,基于Sa VRNN的层次化结构,本文将多层似然进行融合以改进传统的异常检测策略.在两个公共数据集和一个实测数据集上显示所提方法在F1-score指标上实现了相比目前主流方法的更优性能,证明了所提模型在时间序列在线异常检测任务上的有效性.
To characterize the complex temporal dependence and stochasticity of multivariate time series(MTS)and achieve their unsupervised anomaly detection,which is crucial for managing service quality in industrial devices and Internet infrastructures.Our proposed method,called stacked adversarial variational recurrent neural network(SaVRNN),is a hierarchical probabilistic dynamical model that unifies various elements,including hierarchical generative models,stacked recurrent structure,and multilayer adversarial optimization within a Bayesian framework.The fundamental concept behind SaVRNN is to capture the typical patterns in input time series data by considering multilevel and long-range temporal dependencies through stacked recurrent structures and multilayer shape characteristics via hierarchical probabilistic generative operations.Subsequently,we utilize reconstruction probabilities to identify anomalies.To facilitate efficient inference,we introduce a unique upward-downward adversarial inference scheme,which accurately approximates latent variables'posterior distribution.However,given the challenges associated with balancing the training of the generator and discriminator in a multilayer adversarial network,SaVRNN can be challenging to train effectively.To address this issue,we have developed an optimization method based on conditional transport.Leveraging the hierarchical probabilistic structure inherent to SaVRNN,we propose a modified anomaly detection metric that combines the likelihood information from multiple layers.We evaluate the performance of SaVRNN through experiments conducted on two publicly available datasets and one real-world dataset.The results demonstrate the efficacy of SaVRNN for online anomaly detection.
作者
陈文超
方博为
代良
陈渤
刘畅
赵小楠
Wenchao CHEN;Bowei FANG;Liang DAI;Bo CHEN;Chang LIU;Xiaonan ZHAO(National Laboratory of Radar Signal Processing,Xidian University,Xi'an 710071,China;Institute of Information Engineering,Chinese Academy of Sciences,University of Chinese Academy of Sciences,Beijing 100093,China;Shandong New Generation Information Industry Technology Research Institute,Ltd.,Jinan 250i00,China)
出处
《中国科学：信息科学》
CSCD
北大核心
2023年第9期1750-1767,共18页
Scientia Sinica(Informationis)
基金
国家自然科学基金(批准号:6220010437,U21B2006,61771361)
雷达信号处理全国重点实验室基金(批准号:JKW202X0X)
陕西省高校青年创新团队支持项目
111引智基地项目(批准号:B18039)
中组部高层次海外人才引进计划资助。
关键词
异常检测
多维时间序列
对抗变分贝叶斯
变分循环神经网络
条件传输
anomaly detection
multivariate time series
adversarial variational Bayesian
variational RNN
conditional transport
