Abstract:Multiple variable time series anomaly detection plays a significant role in fields such as AIOps and intelligent healthcare. However, time series often lack labels, and anomalies are infrequent compared to normal sequences. Moreover, high-dimensional nonlinear correlations and substantial distribution variations among variables exist in multivariate time series. These challenges make it difficult for the model to accurately model time series. While contemporary reconstruction-based models address challenges of data scarcity and time series modeling through GAN frameworks, little attention has been given to mitigating the substantial differences in variable distributions among multiple time series. This paper proposes the Variable-wise Generative Adversarial Transformer (VGAT) for achieving unsupervised anomaly detection, with WGAN as the training framework and Transformer as the fundamental structure for the Duplicator and Discriminator. Firstly, VGAT makes the Generator act as the Duplicator, reconstructing samples through direct replication, simplifying the model structure, and enhancing efficiency. Secondly, VGAT employs the Soft-DTW regularization to facilitate a more authentic replication of normal sequences by the Duplicator. Most importantly, VGAT introduces the Variable-wise Gate, dynamically calculating a gate for each variable in the multivariate time series to balance the differences in variable distributions. After VGAT generates anomaly scores, the SPOT algorithm based on Extreme Value Theory (EVT) automatically computes the threshold, enhancing the adaptability to different data of the model. VGAT demonstrates state-of-the-art (SOTA) performance, surpassing baselines on public datasets. Notably, as the dimensionality of the series increases, VGAT exhibits more significant improvements. On WADI dataset, VGAT achieves an impressive 68.19% enhancement in F1-score, showcasing remarkable progress compared to the current SOTA model.

Unsupervised Anomaly Detection in Multivariate Time Series Through Transformer-based Variational Autoencoder

Variational transformer-based anomaly detection approach for multivariate time series

Time Series Anomaly Detection with a Transformer Residual Autoencoder-Decoder.

Transformer-based multivariate time series anomaly detection using inter-variable attention mechanism

Variable-wise generative adversarial transformer in multivariate time series anomaly detection

A two-stage adversarial Transformer based approach for multivariate industrial time series anomaly detection

Disentangled Anomaly Detection for Multivariate Time Series

Multivariate time series anomaly detection with variational autoencoder and spatial–temporal graph network

LGAT: A Novel Model for Multivariate Time Series Anomaly Detection with Imporved Anomaly Transformer and Learning Graph Structures

Multivariate Time Series Anomaly Detection with Fourier Time Series Transformer

SiET: Spatial information enhanced transformer for multivariate time series anomaly detection

NVAE-GAN Based Approach for Unsupervised Time Series Anomaly Detection

Unsupervised Anomaly Detection Using Variational Auto-Encoder based Feature Extraction

Graph Transformer Network Incorporating Sparse Representation for Multivariate Time Series Anomaly Detection

TransAnomaly: Video Anomaly Detection Using Video Vision Transformer

VELC: A New Variational AutoEncoder Based Model for Time Series Anomaly Detection

Anomaly Transformer: Time Series Anomaly Detection with Association Discrepancy

From anomaly detection to classification with graph attention and transformer for multivariate time series

EdgeConvFormer: Dynamic Graph CNN and Transformer based Anomaly Detection in Multivariate Time Series

DTAAD: Dual Tcn-attention networks for anomaly detection in multivariate time series data

An Unsupervised Method for Industrial Image Anomaly Detection with Vision Transformer-Based Autoencoder