Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

ALGAN: Time Series Anomaly Detection with Adjusted-LSTM GAN

Md Abul Bashar, Richi Nayak

TL;DR

ALGAN addresses unsupervised anomaly detection in time series by combining an Adjusted-LSTM (ALstm) with a GAN to learn the normal data distribution and detect deviations via a latent-space reconstruction framework. The model introduces two attention layers that refine both LSTM outputs and inputs, and uses a reconstruction-based anomaly score blending residual and discriminator-derived terms. Empirical results on 46 NAB univariate datasets and a large SWaT multivariate dataset show ALGAN consistently outperforms traditional, neural-network, and other GAN-based methods, with particularly strong F1 and recall. The work demonstrates the effectiveness of combining attention-enhanced LSTM units with adversarial learning to capture complex temporal dynamics and detect subtle anomalies in diverse domains, while noting challenges in window-length selection and GAN training stability.

Abstract

Anomaly detection in time series data, to identify points that deviate from normal behaviour, is a common problem in various domains such as manufacturing, medical imaging, and cybersecurity. Recently, Generative Adversarial Networks (GANs) are shown to be effective in detecting anomalies in time series data. The neural network architecture of GANs (i.e. Generator and Discriminator) can significantly improve anomaly detection accuracy. In this paper, we propose a new GAN model, named Adjusted-LSTM GAN (ALGAN), which adjusts the output of an LSTM network for improved anomaly detection in both univariate and multivariate time series data in an unsupervised setting. We evaluate the performance of ALGAN on 46 real-world univariate time series datasets and a large multivariate dataset that spans multiple domains. Our experiments demonstrate that ALGAN outperforms traditional, neural network-based, and other GAN-based methods for anomaly detection in time series data.

ALGAN: Time Series Anomaly Detection with Adjusted-LSTM GAN

TL;DR

ALGAN addresses unsupervised anomaly detection in time series by combining an Adjusted-LSTM (ALstm) with a GAN to learn the normal data distribution and detect deviations via a latent-space reconstruction framework. The model introduces two attention layers that refine both LSTM outputs and inputs, and uses a reconstruction-based anomaly score blending residual and discriminator-derived terms. Empirical results on 46 NAB univariate datasets and a large SWaT multivariate dataset show ALGAN consistently outperforms traditional, neural-network, and other GAN-based methods, with particularly strong F1 and recall. The work demonstrates the effectiveness of combining attention-enhanced LSTM units with adversarial learning to capture complex temporal dynamics and detect subtle anomalies in diverse domains, while noting challenges in window-length selection and GAN training stability.

Abstract

Anomaly detection in time series data, to identify points that deviate from normal behaviour, is a common problem in various domains such as manufacturing, medical imaging, and cybersecurity. Recently, Generative Adversarial Networks (GANs) are shown to be effective in detecting anomalies in time series data. The neural network architecture of GANs (i.e. Generator and Discriminator) can significantly improve anomaly detection accuracy. In this paper, we propose a new GAN model, named Adjusted-LSTM GAN (ALGAN), which adjusts the output of an LSTM network for improved anomaly detection in both univariate and multivariate time series data in an unsupervised setting. We evaluate the performance of ALGAN on 46 real-world univariate time series datasets and a large multivariate dataset that spans multiple domains. Our experiments demonstrate that ALGAN outperforms traditional, neural network-based, and other GAN-based methods for anomaly detection in time series data.
Paper Structure (20 sections, 10 equations, 8 figures, 2 tables, 1 algorithm)

This paper contains 20 sections, 10 equations, 8 figures, 2 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Work
  3. The Proposed Adjusted LSTM Model
  4. Internal Mechanisms of LSTM
  5. Internal Mechanisms of Self-Attention
  6. Mechanisms of the proposed Adjusted LSTM
  7. Generative Adversarial Representation Learning to Identify Time Series Anomalies
  8. Learning General Data Distribution
  9. Mapping Real-Data to the Latent Space
  10. Detection of Anomalies
  11. Empirical Evaluation
  12. Data Collection
  13. NAB Dataset Collection
  14. SWaT Dataset
  15. Evaluation Measures
  16. ...and 5 more sections

Figures (8)

  • Figure 1: Weight Adjusted LSTM suited to handle Time Series data
  • Figure 2: ALGAN: Time Series Anomaly Detection with Adjusted-LSTM GAN
  • Figure 3: Generator and Discriminator Architecture employed in ALGAN
  • Figure 4: Visualisation of Anomalies Detection by ALGAN. Note the change in scaling in the Y axis of (a) and (b); as well as (C) and (d)
  • Figure 5: Cumulative ranking obtained by individually ranking each model per dataset per measure
  • ...and 3 more figures