基于多特征融合的应用系统监控指标异常检测方法  

ANOMALY DETECTION METHOD FOR KPIS IN APPLICATION SYSTEMS BASED ON MULTI-FEATURE FUSION

作　　者：曹钰聪 张俊[1] Cao Yucong;Zhang Jun(Information Science and Technology College,Dalian Maritime University,Dalian 116026,Liaoning,China)

机构地区：[1]大连海事大学信息科学技术学院,辽宁大连116026

出　　处：《计算机应用与软件》2024年第8期74-83,共10页Computer Applications and Software

基　　金：国家自然科学基金项目(61976032)。

摘　　要：为解决现有监控指标异常检测技术存在的特征学习不充分、阈值固定等问题,提出一种基于多特征融合的应用系统监控指标异常检测方法。使用1D-CNN(1D-Convolutional Neural Network)与SRNN(Stochastic Recurrent Neural Network)提取数据特征,引入SE块(Squeeze-and-Excitation)突出指标关键特征以优化特征提取,加强分类效果。以VAE(Variational Auto-Encoder)为框架计算数据重构概率,并通过优化的极值模型计算最优异常阈值以判断异常。实验结果表明,所提方法在基于两个公开数据集的异常检测任务的F1评分最优达到92%,优于目前先进的异常检测方法。In order to solve the problems of existing KPIs anomaly detection methods,such as insufficient feature learning and fixed thresholds,we propose a anomaly detection method for KPIs in application systems based on multi-feature fusion.We used the 1D-convolutional neural network(1D-CNN)and stochastic recurrent neural network(SRNN)to extract data features,and introduced the squeeze-and-excitation(SE)block to highlight the key features of KPIs to optimize feature extraction and strengthen the classification effect.We used the variational auto-encoder(VAE)as the framework to calculate the reconstruction probability of data,and calculated the best anomaly threshold through the extreme value model to determine anomalies.Experimental results show that the proposed method can effectively detect outlier on two public datasets,with best F1 score of 92%,and has better performance than some advanced anomaly detection methods.

关 键 词：监控指标 异常检测 特征提取 变分自编码器 极值理论 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]