TCAD: Unsupervised Anomaly Detection Based on Global Local Representation Differences

Multivariate time series anomaly detection is of great interest because of its wide range of applications. Since it is difficult to obtain accurate anomaly labels, many unsupervised anomaly detection algorithms have been developed. However, it is challenging to build an unsupervised multivariate ano...

Full description

Saved in:
Bibliographic Details
Published in:IEEE access 2022, Vol.10, p.114683-114693
Main Authors: Duan, Yaqiong, Xiang, Mian, Zhou, Bingtao, Fu, Desu, Liu, Hongxiao
Format: Article
Language:English
Subjects:
Algorithms
Anomalies
Anomaly detection
Datasets
Feature extraction
Image reconstruction
Learning systems
Multivariate analysis
multivariate time series
Representations
Resnet
Sequences
Time series analysis
transformer
Transformers
Unsupervised Learning
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Multivariate time series anomaly detection is of great interest because of its wide range of applications. Since it is difficult to obtain accurate anomaly labels, many unsupervised anomaly detection algorithms have been developed. However, it is challenging to build an unsupervised multivariate anomaly detection model because we need to find a criterion with anomaly discriminative power. Previously, researchers have focused on extracting the association of time points with global sequences, while ignoring the association of time points with local sequences. In this paper, we propose a combined model TCAD based on Transformer and Resnet, which learns global and local features of sequences using Transformer and Resnet, and constrains the learning of rich global local representations using reconstructed differences and global local representation differences. In addition, this paper proposes an anomaly score based on global and local feature discrepancies. TCAD is extensively tested on four public datasets and two private datasets.
ISSN:2169-3536
2169-3536
DOI:10.1109/ACCESS.2022.3216930