基于深度自编码器的自适应异常检测算法及其应用研究  

Research on deep autoencoder-based adaptive anomaly detection algorithm and its application

作　　者：陈晓红[1,2,3] 陈姣龙 胡东滨 梁伟[2,3] 张威威 CHEN Xiaohong;CHEN Jiaolong;HU Dongbin;LIANG Wei;ZHANG Weiwei(School of Business,Central South University,Changsha 410083,China;Xiangjiang Laboratory,Changsha 410205,China;School of Frontier Crossover Studies,Hunan University of Technology and Business,Changsha 410205,China)

机构地区：[1]中南大学商学院,长沙410083 [2]湘江实验室,长沙410205 [3]湖南工商大学前沿交叉学院,长沙410205

出　　处：《系统工程理论与实践》2024年第8期2718-2732,共15页Systems Engineering-Theory & Practice

基　　金：国家自然科学基金基础科学中心项目(72088101);湘江实验室重大项目(23XJ01006);国家自然科学基金重大项目(71790615)。

摘　　要：随着深度学习的发展,深度自编码器被广泛应用于异常检测任务中.然而,现有一些基于深度自编码器的异常检测算法仍存在数据分布复杂多样、误报率、漏报率高等问题.为解决以上问题,本文提出了一种基于深度自编码器的自适应异常检测算法,该算法采用基于密度峰值的自适应地标过滤机制,自适应地选择密度高的样本作为候选地标中心,旨在发现潜在特征空间中正常样本的多样性.其次,利用地标过滤机制对候选地标中心进行过滤和优化,以此增强地标中心的代表性和稀疏性.然后,进一步设计了一种新颖的损失函数来迭代优化模型参数,旨在增强正常样本与地标中心之间的相关性.最后,将所提出的异常检测算法应用于电池故障诊断,实证结果表明本文的工作在准确度、误报率和漏报率等方面显著优于现有一些异常检测算法,可以有效识别故障电池,能够为电池故障识别与状态管理提供技术支持和精准服务.With the development of deep learning,deep autoencoder has been widely applied in anomaly detection with efficient data encoding and reconstruction mechanisms.However,some existing deep autoencoder-based anomaly detection algorithms still face many problems,such as complex and diverse data distributions,high false alarm rate,and high missing alarm rate,etc.To overcome the above-mentioned problems,we propose a deep autoencoder-based adaptive anomaly detection algorithm.The algorithm utilizes an adaptive landmark filtering mechanism via density peak,which can select some normal samples with high density as candidate landmarks,aiming to discover the diversity of normal samples in the latent feature space.Subsequently,the landmark filtering mechanism is employed to filter and optimize the candidate landmarks to enhance the representativeness and sparsity of the landmarks.Furthermore,we design a novel loss function to optimize the model parameters iteratively with the aim of enhancing the correlation between normal samples and landmarks.Finally,the proposed anomaly detection algorithm is applied to a battery fault diagnosis,and the experiment results demonstrate that this work outperforms the existing anomaly detection algorithms in terms of accuracy,false alarm rate,and missing alarm rate.It can identify faulty batteries effectively,and provide the technical support and precise services for the battery fault identification and state management.

关 键 词：异常检测 深度自编码器 自适应地标过滤机制 动力电池 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]