基于鲁棒变分自动编码器的时序异常检测  

Time series anomaly detection based on robust variational autoencoder

作　　者：冯志鹏 赵旭俊[1] FENG Zhi-peng;ZHAO Xu-jun(College of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China)

机构地区：[1]太原科技大学计算机科学与技术学院,山西太原030024

出　　处：《计算机工程与设计》2025年第2期376-383,共8页Computer Engineering and Design

基　　金：国家自然科学基金项目(61572343);山西省基础研究计划基金项目(202303021221142)。

摘　　要：针对变分自动编码器在时序异常检测中鲁棒性低的问题,提出一种鲁棒变分自动编码器异常检测算法。通过对时间信息进行编码并融合在变分自动编码器的隐藏层中,有效学习不同时间窗口间的周期性模式;添加Dropout层以防止过拟合,增强局部特征的学习,增强时序关系在异常检测中的作用。提出一种损失函数来识别异常时序数据,通过重构误差与阈值的比较筛选异常。利用交替方向乘子法对异常时序进行验证,提高鲁棒性。在4个真实数据集上与4种基准方法相比,该算法在异常样本上的精度和F1分数均有显著提升。A robust variational autoencoder anomaly detection algorithm was proposed to address the issue of low robustness in timing anomaly detection.By encoding time information and fusing it into the hidden layer of the variational autoencoder,periodic patterns between different time windows were effectively learned,and a Dropout layer was added to prevent overfitting,which enhanced the learning of local features and enhanced the role of temporal relationships in anomaly detection.A loss function was proposed to identify anomalous temporal data,and anomalies were filtered by comparing the reconstruction error with the thres-hold.And the alternating direction multiplier method was used to verify abnormal time series and improve the robustness.Compared with the four benchmark methods on four real datasets,the algorithm shows significant improvements in accuracy and F1 score on anomalous samples.

关 键 词：时间序列 自动编码器 神经网络 鲁棒 损失函数 重构误差 异常检测 

分 类 号：TP311[自动化与计算机技术—计算机软件与理论]