[Paper Review] Deep Learning for Time Series Anomaly Detection: A Survey

Zahra Zamanzadeh Darban, Geoffrey I. Webb+3 more|arXiv (Cornell University)|Nov 9, 2022

Anomaly Detection Techniques and Applications182 references74 citations

TL;DR

This survey provides a comprehensive taxonomy and state-of-the-art overview of deep learning methods for time series anomaly detection, comparing forecasting-, reconstruction-, and hybrid-based approaches, datasets, and evaluation practices.

ABSTRACT

Time series anomaly detection has applications in a wide range of research fields and applications, including manufacturing and healthcare. The presence of anomalies can indicate novel or unexpected events, such as production faults, system defects, or heart fluttering, and is therefore of particular interest. The large size and complex patterns of time series have led researchers to develop specialised deep learning models for detecting anomalous patterns. This survey focuses on providing structured and comprehensive state-of-the-art time series anomaly detection models through the use of deep learning. It providing a taxonomy based on the factors that divide anomaly detection models into different categories. Aside from describing the basic anomaly detection technique for each category, the advantages and limitations are also discussed. Furthermore, this study includes examples of deep anomaly detection in time series across various application domains in recent years. It finally summarises open issues in research and challenges faced while adopting deep anomaly detection models.

Motivation & Objective

• Propose a novel taxonomy of deep anomaly detection models for time series, categorized into forecasting-based, reconstruction-based, and hybrid methods.
• Review current state-of-the-art models, architectures, and learning schemes used for univariate and multivariate time series anomaly detection.
• Identify primary benchmarks, datasets, and practical evaluation protocols used in the field.
• Discuss fundamental principles behind different anomaly types in time series and highlight challenges and future research directions.

Proposed method

• Classify deep time series anomaly detection models into forecasting-based, reconstruction-based, and hybrid categories.
• Further subdivide models by main architectures (e.g., RNN, LSTM, CNN, GNN, GAN, VAE, Transformer) and input type (univariate vs multivariate).
• Explain learning schemes (supervised, unsupervised, semi-supervised, self-supervised) and how anomaly scores are derived (prediction error, reconstruction probability).
• Describe input processing (point vs subsequence, sliding windows) and output forms (anomaly scores vs binary labels).
• Discuss interpretability measures and evaluation protocols, including thresholding strategies and point-adjustment considerations.
• Summarize publicly available datasets and provide a structured comparison framework (Tables 1 & 2 in the source).

Experimental results

Research questions

• RQ1 What is the comprehensive taxonomy of deep learning models for time series anomaly detection and their architectural characteristics?
• RQ2 What are the main datasets and benchmarks used to evaluate deep time series anomaly detection methods?
• RQ3 What evaluation protocols and thresholds are used to determine anomalies in time series data, and how do interpretability considerations influence practice?
• RQ4 What open issues, challenges, and future directions exist for applying deep anomaly detection to time series data?

Key findings

• A novel taxonomy is proposed, dividing models into forecasting-based, reconstruction-based, and hybrid approaches, with subcategories by neural architectures.
• A comprehensive review of state-of-the-art models and architectures (e.g., RNN, LSTM, CNN, GNN, GAN, VAE, Transformer) across univariate and multivariate time series is provided.
• The survey collates primary benchmarks and datasets, detailing commonly used sources and hyperlinks for researchers.
• The paper discusses learning schemes (supervised, unsupervised, semi-supervised, self-supervised) and how anomaly scores are typically derived from prediction errors or reconstruction probabilities.
• Evaluation practices such as point adjustment (PA) and PA%K protocols are discussed, highlighting potential biases in F1 scores without rigorous evaluation.
• Open issues and challenges in adopting deep anomaly detection for time series data are identified, outlining directions for future research.