基于 AAE 的网络性能异常发现-《计算机技术与发展》

文章信息/Info

Author(s):: WANG Tian-feng; HU Gu-yu; WANG Rui; PENG Dong-yang; School of Command and Control Engineering,Army Engineering University of PLA,Nanjing 210007,China

Keywords:: network operation and maintenance; temporal sequence; anomaly detection; AAE; GAN; KNN

摘要:: 网络运维在充分发挥网络潜能方面有着不可替代的作用。其中,网络关键性能的监测和维护尤为重要。使用智能化的方法自动发现网络 KPI 的异常能够极大地减少运维人工成本,提升网络运维的效率。人工方法标注网络 KPI 中的异常,难度高,耗时长,因此无监督学习的异常检测正在成为解决此类问题的主要方法。提出一种基于对抗自编码器 AAE的无监督检测模型 AAE-AD,可以自动发现网络 KPI 中出现的异常,以便分析和排除网络故障。 AAE-AD 中使用了 K 最近邻算法进行缺失值的填充,交替训练自编码器网络和鉴别器网络来捕获正常数据的分布模式,结合自编码器网络的重构误差和鉴别器网络的鉴别能力计算出异常分值。实验表明,AAE-AD 模型在最优 F 分数指标和 AUC 指标上均优于其他的无监督异常检测模型。

Abstract:: Network operation and maintenance plays an irreplaceable role in giving full play to the potential of network. The monitoring and maintenance of network critical performance is particularly important. Automatic detection of anomalies in network KPI by intelligent methods can greatly reduce the labor cost and improve the? efficiency of network operation and maintenance. Manual methods to an notate anomalies in network KPIs are difficult and time-consuming,so unsupervised learning anomaly detection is becoming the main method to solve such problems. We propose an unsupervised detection model,AAE-AD,which is based on the adversarial autoencoder AAE,whichcan automatically detect the anomalies in network KPI,so as to analyze and troubleshoot network faults. In AAE-AD,the missing valueis filled by the KNN,and the autoencoder network and the discriminator network are trained alternatively to capture the distribution pattern of normal data. The abnormal score is calculated by combining the reconstruction error of the autoencoder network and the discriminator network’s discrimination ability. The experiment shows that the AAE-AD model is superior to other unsupervised anomaly detection models in terms of the optimal F score and AUC.