一种采用串行自编码器的时序数据异常检测方法  

A serial autoencoders based method for detecting time series anomalies

作　　者：徐天慧 郭强[1,2,3] 张彩明[2,3,4] XU Tian-hui;GUO Qiang;ZHANG Cai-ming(School of Computer Science and Technology,Shandong University of Finance and Economics,Jinan 250014,China;Shandong Key Laboratory of Digital Media Technology,Shandong University of Finance and Economics,Jinan 250014,China;Shandong Provincial Laboratory of Future Intelligence and Financial Engineering,Yantai 264005,China;Software College,Shandong University,Jinan 250101,China)

机构地区：[1]山东财经大学计算机科学与技术学院,济南250014 [2]山东财经大学山东省数字媒体技术重点实验室,济南250014 [3]山东省未来智能金融工程实验室,山东烟台264005 [4]山东大学软件学院,济南250101

出　　处：《控制与决策》2023年第12期3507-3515,共9页Control and Decision

基　　金：国家自然科学基金项目(61873145,61802229);山东省自然科学省属高校优秀青年人才联合基金项目(ZR2017JL029);山东省高等学校青创科技支持计划项目(2019KJN045)。

摘　　要：基于深度学习的时序数据异常检测模型大多采用循环神经网络或长短期记忆网络捕捉时序依赖性,并利用自编码器重构数据,进而实现时序数据的异常检测.虽然此类检测模型实现了较高的异常检测率,但它们的网络结构复杂,导致模型的计算效率较低.为提高模型的计算效率,提出一种基于串行自编码器的异常检测模型SAE-AD.该模型仅包含两个结构简单的自编码器(AE_(1)和AE_(2)),其所含参数较少,且训练目标较为简单,从而加快了模型的计算效率.通过将自编码器AE_(1)和AE_(2)串行拼接,即AE_(1)的输出作为AE_(2)的输入,可有效提高AE_(2)的解码器对正常数据特征的解码能力,有助于提升模型的检测准确率.实验结果表明,相较于其他新近提出的异常检测模型,SAE-AD模型具有更高的精确率、召回率和F_(1)值.Aiming to detect time series anomalies,deep learning methods generally use the recurrent neural network or long short term memory to capture temporal dependency,and adopt autoencoder to reconstruct data.Although they work well for detecting anomalies,the network structures of these methods are complex,resulting in slow computational efficiency.In order to improve the computational efficiency,this paper proposes a method called serial autoencoders based anomaly detection (SAE-AD) which contains two autoencoders AE_(1) and AE_(2) with simple structure.Due to the simplicity,there are a few training parameters and its training objectiv function is relatively simple,which speeds up the computation.In addition,the output of AE_(1) is fed into AE_(2)to improve the decoding ability of the decoder of AE_(2)This way of serial training makes SAE-AD achieve better detection accuracy.Experiment results show that the proposed method has better precision,recall,F_(1) score than several state-of-art anomaly detection methods.

关 键 词：深度学习 时序数据 异常检测 自编码器 数据重构 编码器 解码器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]