«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

期数:

2025年07期

页码:

108-116

栏目:

人工智能

出版日期:

2025-07-10

文章信息/Info

Title:

Multivariate Time Series Anomaly Detection Based on AE-STCN

文章编号:

1673-629X(2025)07-0108-09

作者:

洪培林1; 曾碧卿2; 刘馨瑶1

1. 华南师范大学 人工智能学院,广东 佛山 528225;

2. 华南师范大学 阿伯丁数据科学与人工智能学院,广东 佛山 528225

Author(s):

HONG Pei-lin1; ZENG Bi-qing2; LIU Xin-yao1

1. School of Artificial Intelligence,South China Normal University,Foshan 528225,China;

2. Aberdeen Institute of Data Science and Artificial Intelligence,South China Normal University,Foshan 528225,China

关键词:

多元时间序列; 异常检测; 滤波器; 自编码器; 时域卷积; 联合优化

Keywords:

multivariate time series; anomaly detection; filter; auto-encoder; temporal convolution network; joint optimization

分类号:

TP18

DOI:

10.20165/j.cnki.ISSN1673-629X.2025.0037

摘要:

随着工业领域中传感器技术的发展,数据量和复杂性迅速增加,传统的异常检测方法在面对噪声干扰和复杂数据模式时,显得力不从心。 该文提出一种多元时序异常检测模型 AE-STCN,结合了自动编码器(AE)和对称时域卷积网络(STCN)的优势,联合使用预测与重构的方法进行优化,以更准确地捕捉异常模式。 其中自动编码器通过学习数据的内在结构进行时间序列重构,改进的对称时域卷积网络对输入序列进行镜像翻转预测未来时刻的值,进一步捕捉时间序列中的时序依赖性。考虑到训练数据中的噪声污染问题,该文还提出了一种基于 Transformer 的滤波器模块,有效减弱了噪声对模型训练的负面影响,增强了对正常模式的学习能力。为验证模型的性能,在 3 个公开数据集对 AE-STCN 进行了实验,并在 AUC、 F1 和 Fc1 指标的综合评估下实现了最佳性能。 结果表明 AE-STCN 优于所有基线模型,充分证明了 AE-STCN 在处理多元时间序列数据时的有效性和优越性,为多元时序异常检测提供了一种新的、可靠的解决方案。

Abstract:

With the development of sensor technology in the industrial field,the amount and complexity of data have increased rapidly.Traditional anomaly detection methods seem to be inadequate when faced with noise interference and complex data patterns. We propose a multivariate time series anomaly detection model named AE - STCN, which combines the advantages of AutoEncoder ( AE) and Symmetric Temporal Convolutional Network (STCN) and jointly uses prediction and reconstruction methods for optimization to capture abnormal patterns more accurately. Among them,the AutoEncoder reconstructs time series by learning the internal structure of data,and the improved Symmetric Temporal Convolutional Network mirrors and flips the input sequence to predict the values at future moments, further capturing the temporal dependencies in the time series. Considering the problem of noise pollution in the training data,we also propose a Transformer - based filter module to effectively reduce the negative influence of noise on model training and enhance the learning ability of normal patterns. To verify the performance of the model,we conduct experiments on the AE-STCN in three public datasets and achieve the best performance under the comprehensive evaluation of AUC, F1 ,and Fc1 metrics. The results show that AE-STCN outperforms all the baseline models,fully demonstrating the effectiveness and superiority of AE-STCN in processing multivariate time series data and providing a new and reliable solution for multivariate time series anomaly detection.

常用功能

更新日期/Last Update: 2025-07-10