连续无监督异常检测  

Continual unsupervised anomaly detection

作　　者：倪一鸣 陈松灿[1,2] Yiming NI;Songcan CHEN(College of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 210016,China;MIIT Key Laboratory of Pattern Analysis and Machine Intelligence,Nanjing 210016,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京210016 [2]模式分析与机器智能工业和信息化部重点实验室,南京210016

出　　处：《中国科学：信息科学》2022年第1期75-85,共11页Scientia Sinica(Informationis)

基　　金：航空发动机及燃气轮机重大专项基础研究项目(批准号:J2019-IV-0018-0086);国家自然科学基金(批准号:62076124)资助项目。

摘　　要：无监督异常检测(unsupervised anomaly detection,UAD)旨在检测任何未见过的偏离预期模式或正常分布的数据,由于其学习过程不依赖对罕见异常样本的获取,因此在现实动态环境下备受青睐.然而,在现实场景中,目标任务往往会随时间动态变化,这要求模型能够连续执行多个不同的UAD任务,确保在仅有当前任务正常数据的前提下,实现对所有见过任务的异常检测.本文旨在研究这一问题,尝试从互信息角度,提出一种新的连续UAD(CUAD)算法.具体而言,我们针对原始目标依赖过往任务原始数据和异常数据的问题,给出基于信息论的损失函数,并对其进行近似优化.据此,我们构建出来的深度编码器模型既能连续执行不同的UAD任务,又能有效应对连续学习带来的灾难性遗忘问题.最后,我们在多个标准数据集上的实验验证了所提出方法的优越性.Unsupervised anomaly detection(UAD)involves detecting unseen data that differs from expected or normal patterns.UAD has gained extensive attention in dynamic scenarios because its learning process is independent of rare abnormal samples.In reality,target tasks change dynamically with time,so a model must solve different UAD problems continuously and effectively,detecting all kinds of previous anomalies with only normal data in the current task.Thus,we propose a novel continual UAD(CUAD)algorithm based on mutual information.Specifically,the original objective function needs previous and abnormal data,which are missing.To deal with this problem,we introduce an objective function and then approximate and optimize it based on information theory.On the basis of this,a deep encoder is used to continuously detect various anomalies while efficiently alleviating catastrophic forgetting caused by continual learning.Experiments on several datasets demonstrate that the proposed model outperforms state-of-the-art methods.

关 键 词：异常检测 无监督 灾难性遗忘 连续学习 信息论 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TP18[自动化与计算机技术—计算机科学与技术]