基于残差时序结构建模的时间序列异常检测框架
RTS-AD: A Residual Temporal Structure Modeling Framework for Time Series Anomaly Detection
王志娴 1周文安1
作者信息
- 1. 北京邮电大学计算机学院，北京100876
- 折叠
摘要
移动网络流量异常检测已成为支撑网络智能运维的关键任务。然而,受业务需求、用户行为变化等因素的共同影响,真实流量序列通常具有显著的复杂性与非平稳性,使得异常检测十分困难。现有的深度学习方法通常利用残差幅值提取异常证据：基于预测的方法通过单点预测误差刻画观测相对正常行为的瞬时偏离,基于重构的方法通过重构困难程度反映异常对正常表示结构的破坏,二者具有一定互补性。然而,无论是单独还是组合使用这些残差,现有方法大多仍主要依据其幅值进行判别,对残差序列内部的时序结构利用不足。针对上述问题,本文提出一种基于残差时序结构建模的时间序列异常检测框架 RTS-AD,通过残差建模模块对标准化残差序列进行显式建模,以提升异常识别能力。该模块由概率残差生成器与残差模式重构器组成：前者通过概率预测条件均值与预测不确定性构造标准化残差序列,后者在该残差空间中学习正常残差模式的时序结构。在 NAB 真实异常数据集上的实验结果表明,RTS-AD 在多项检测指标上优于多种强基线方法,验证了方法的有效性。
Abstract
Mobile network traffic anomaly detection has become a key task for supporting intelligent network operation and maintenance. However, under the joint influence of service demand variation, user behavior evolution, and external disturbances, real-world traffic series usually exhibit significant complexity and non-stationarity, making anomaly detection highly challenging. Existing deep learning methods typically extract anomaly evidence from residual magnitudes: prediction-based methods characterize the instantaneous deviation of observations from normal behavior through point-wise prediction errors, while reconstruction-based methods reflect the disruption of normal representation structures through reconstruction difficulty; these two paradigms are therefore complementary to some extent. Nevertheless, whether these residuals are used individually or in combination, most existing methods still rely primarily on their magnitudes for anomaly discrimination and make insufficient use of the temporal structure within residual sequences. To address this issue, this paper proposes RTS-AD, a residual temporal structure modeling framework for time series anomaly detection, which explicitly models standardized residual sequences through a residual modeling module to improve anomaly identification. This module consists of a probabilistic residual generator and a residual pattern reconstructor: the former constructs standardized residual sequences by probabilistically predicting the conditional mean and predictive uncertainty, while the latter learns the temporal structure of normal residual patterns in the residual space. Experimental results on the NAB real-world anomaly dataset show that RTS-AD outperforms multiple strong baselines on several detection metrics, demonstrating the effectiveness of the proposed method.关键词
计算机应用技术/时间序列异常检测/概率预测/移动网络流量Key words
time series anomaly detection/probabilistic forecasting/mobile network traffic引用本文复制引用
王志娴,周文安.基于残差时序结构建模的时间序列异常检测框架[EB/OL].(2026-04-08)[2026-04-11].http://www.paper.edu.cn/releasepaper/content/202604-70.学科分类
计算技术、计算机技术
评论