Unified Taxonomy for Multivariate Time Series Anomaly Detection using Deep Learning

Bruna Alves; Armando J. Pinho; Sónia Gouveia

stat.MLMar 19, 2026

Unified Taxonomy for Multivariate Time Series Anomaly Detection using Deep Learning

Authors: Bruna Alves, Armando J. Pinho, Sónia Gouveia

Organizations: IEETA/DETI/LASI, University of Aveiro, 3810-193, Aveiro, Portugal.

Abstract

The topic of Multivariate Time Series Anomaly Detection (MTSAD) has grown rapidly over the past years, with a steady rise in publications and Deep Learning (DL) models becoming the dominant paradigm. To address the lack of systematization in the field, this study introduces a novel and unified taxonomy with eleven dimensions over three parts (Input, Output and Model) for the categorization of DL-based MTSAD methods. The dimensions were established in a two-fold approach. First, they derived from a comprehensive analysis of methodological studies. Second, insights from review papers were incorporated. Furthermore, the proposed taxonomy was validated using an additional set of recent publications, providing a clear overview of methodological trends in MTSAD. Results reveal a convergence toward Transformer-based and reconstruction and prediction models, setting the foundation for emerging adaptive and generative trends. Building on and complementing existing surveys, this unified taxonomy is designed to accommodate future developments, allowing for new categories or dimensions to be added as the field progresses. This work thus consolidates fragmented knowledge in the field and provides a reference point for future research in MTSAD.

Figures & tables

Figure 1: Examples of temporal anomalies in synthetic MTS generated with https://github.com/datamllab/tods/tree/benchmark/benchmark/synthetic/Generator : (A) Point anomaly in metrics 1 and 2, (B) Contextual in metric 1, (C) Collective in metric 1 and (D) Collective spanning metric 3. The anomaly in (C) is also inter-metric.

Figure 2: Schematic representation of the process of learning a regression model $f$ . In case of a classification model, the output of the model would be $\hat{y}$ and the loss function would compare it with the know target $y$ , $\mathcal{L}(y,\hat{y})$ .

Figure 3: Illustration of the point-adjustment (PA) strategy showing ground truth, model output and adjusted output over time points (circles). Image adapted from ( El Amine Sehili and Zhang, 2024 ) .

Figure 4: Paper screening and selection process. Starting from 898 screened documents (2019–2024), 393 DL-based papers were selected and combined with 16 papers from reference list screening, resulting in 409 categorized papers. Additionally, 12 review papers were identified to complement the proposed taxonomy.

Figure 5: Annual number of documents retrieved using the search query. A total of 1,040 documents were retrieved, including 142 before 2019 and 898 from 2019 to 2024. The red-shaded area highlights the 2019–2024 period, which marks a significant growth in publication volume. Chart directly retrieved and adapted from Scopus ®( https://www.scopus.com/ ).

Figure 6: Distribution of the 898 Scopus documents screened from the 2019-2024 period into eligible (green) and excluded (red) categories. Among the 464 eligible, 393 correspond to DL-based methods, the focus of this review. The 434 excluded were based on six criteria (12, 86, 168, 106, 50, 12).

Figure 7: Schematic diagram of the Model dimensions (in bold). The dashed border indicates that the Feature Extraction module is optional.

Part	Dimension	Categories	Description
Input	Preprocessing	Yes \| No	Raw TS \| Transformed TS
	Format	Point	Single time step $t$ (vector format)
		Window	Consecutive time steps (matrix format)
		Image	Transformed version of a window (matrix format)
		Window + Image	Combination of Window and Image
	Content	Time	Temporal ordering

Table 1: Proposed taxonomy of DL-based approaches for MTSAD, structured into three main parts and several dimensions that reflect key modeling choices.

Part	Dimension	Choi et al. (2021)	Li and Jung (2023)	Darban et al. (2024)	Correia et al. (2024)
Input	Preprocessing	—	—	—	—
	Format	—	—	Point, Window	Sliding window, Input sequence
	Content	—	—	—	—
Output	Type	Reconstruction, Prediction, Dissimilarity	—	Forecasting, Reconstruction, Representation, Hybrid	Prediction, Reconstruction
	Optimal Threshold	—	Fixed, Optimal, Adaptive, Specific, Non-parametric	—	—
	Time-variyng Threshold

Table 2: Literature contributions to the novel of taxonomy proposed in this work. — indicates it is absent or not explicitly defined.

Figure 8: Sunburst chart of 409 categorized documents across the taxonomy. The inner circle shows the Parts, followed by Dimensions and Categories. Colors represent the number of documents and “n.e.” (not explicit) refers to papers that did not explicitly report information for a given dimension. The interactive version of this figure is available at https://ieeta-pt.github.io/DL-MTSAD/Sunburst.html .

Panel of Figure 9 — Figure 9: Schematic diagram representing the novel categories of the T/S dependency dimension, (a) Sequential T $\rightarrow$ S (top) and S $\rightarrow$ T (bottom), (b) Nested and (c) Parallel.

Figure 10: Configuration patterns of models across T/S dependency categories (One-dimensional, Sequential, Nested and Parallel). Analysis excluded the Loss Function, Optimal Threshold, and Time-varying Threshold dimensions.

Time-Series Anomaly Detection

Time-Series Anomaly Detection identifies unusual patterns or events in sequential data, from sensor readings and network traffic to health and industrial telemetry.

See all 203 papers

Future of Time Series Anomaly Detection: From Historical Review to Unified Outlook
Oct 22, 2026Zhijie Zhong, Zhiwen Yu, Pengyu Li +6Time-Series Anomaly Detection Anomaly Detection
School of Future Technology, South China University of Technology, Guangzhou, 510650, China · Pengcheng Laboratory, Shenzhen, 518066, China · School of Computer Science and Engineering, South China University of Technology, Guangzhou, Guangdong 510650, China +3
Benchmarking Inductive Biases for Multivariate Time-Series Anomaly Detection with a Robust Multi-View Channel-Graph Detector
May 27, 2026Junhao Wei, Yanxiao Li, Bidong Chen +9Graph-Based Anomaly Detection Multi-View Learning
Faculty of Applied Sciences, Macao Polytechnic University, Macao SAR, China · 2Macao Polytechnic University, Macao SAR, China
Multivariate Time Series Anomaly Detection via Dual-Branch Reconstruction and Autoregressive Flow-based Residual Density Estimation
Mar 29, 2026Jun Liu, Ying Chen, Ziqian Lu +2Time-Series Anomaly Detection Anomaly Detection
Zhejiang University, Hangzhou, China · Zhejiang Sci-Tech University, Hangzhou, China

Oct 22, 2026J/K move · Enter open · S save

Future of Time Series Anomaly Detection: From Historical Review to Unified Outlook

Zhijie Zhong, Zhiwen Yu, Pengyu Li, Liujian Zhang+5

School of Future Technology, South China University of Technology, Guangzhou, 510650, China · Pengcheng Laboratory, Shenzhen, 518066, China · School of Computer Science and Engineering, South China University of Technology, Guangzhou, Guangdong 510650, China+3

Time-Series Anomaly Detection Anomaly Detection

Open paper AskPDF arXiv

Time series anomaly detection (TSAD) is a foundational problem in data mining, aiming to identify observations that deviate significantly from nominal temporal patterns. Driven by escalating data volumes and application complexity, recent research has yielded diverse architectures, including diffusion models, state space models, and large language models (LLMs). However, existing surveys provide limited coverage of these emerging designs and lack a unified treatment of multivariate anomalies, cross-domain unified modeling, metric reliability, and contemporary evaluation challenges. To bridge these gaps, this survey delivers three core contributions: (1) We establish a refined anomaly categorization and a structured taxonomy of paradigms reflecting modern advances, comparing detection mechanisms and charting developmental trajectories across paradigms. (2) We provide an extensive analysis of evaluation methodologies and conduct comprehensive experiments to assess metric fidelity, complemented by an open-source benchmark repository encompassing algorithms, datasets, and evaluation tools. (3) We outline key open challenges, offering forward-looking perspectives on dynamic thresholding, model interpretability, zero-shot transfer, multi-modal alignment, and foundational models for time series. By addressing existing theoretical and empirical limitations, this survey establishes a coherent framework to guide future TSAD research and practical deployments.

Abstract

Time-Series Anomaly Detection

Explore similar work

Future of Time Series Anomaly Detection: From Historical Review to Unified Outlook

Benchmarking Inductive Biases for Multivariate Time-Series Anomaly Detection with a Robust Multi-View Channel-Graph Detector

Multivariate Time Series Anomaly Detection via Dual-Branch Reconstruction and Autoregressive Flow-based Residual Density Estimation