Research map / Time series / Anomaly detection

VETime: Vision Enhanced Zero-Shot Time Series Anomaly Detection

NeurIPS 2026 · Time series › Anomaly detection · approach: Time series anomaly detection evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes VETime, a zero-shot time-series anomaly detector that fuses a 1D temporal encoder with a frozen MAE vision encoder over reversibly converted images, using patch-level temporal alignment, anomaly-window contrastive learning, and task-adaptive fusion.

Problem

Zero-shot time-series anomaly detection needs both precise pointwise localization of point anomalies and global context detection of context anomalies; 1D temporal foundation models miss global context, while vision-based methods lose temporal alignment and over-detect with coarse windows.

Approach

Converts series into multi-channel (trend/residual/raw) periodic-folded images, aligns ViT features back onto the 1D patch timeline, contrasts visual and temporal features within and across anomaly windows, and fuses modalities with a learned router trained jointly with reconstruction.

Key results

• On 11 univariate TSB-AD datasets, VETime wins 25 of 44 zero-shot first-place slots with average rank 2.05 (Affiliation-F1, F1_T, Standard-F1, VUS-PR).
• Against vision-based baselines on NAB, YAHOO, SMAP, MSL, VETime reaches VUS-PR 32.98 on NAB and 88.61 on YAHOO, and runs at 0.02-0.04 s per series versus roughly 1.7-9.3 s for baselines.
• Ablation: removing Reversible Image Conversion (reverting to line plots) drops NAB VUS-PR from 32.98 to 24.05 and WSD VUS-PR from 33.85 to 18.71.

Surprising findings

• Scaling the vision encoder from MAE-Base to MAE-Large gives only marginal gains and even degrades YAHOO performance despite 3.5x parameters.
• AWCL gives little or negative benefit on densely anomalous YAHOO, where constructing contrastive pairs is hard.

Limitations

• Synthetic training data (0.5B points) generation details are deferred to a cited work; the evaluation depends on strong zero-shot baselines and reported results on a subset of datasets for vision-based comparisons.
• Vision-based comparisons use only four datasets due to token cost, limiting coverage.
• Authors note future work on textual modalities for interpretability.

Experimental setup

Datasets: synthetic anomaly dataset (0.5B points)

Benchmarks: TSB-AD univariate (IOPS, MGAB, NAB, NEK, Power, SED, Stock, TODS, UCR, WSD, YAHOO) TSB-AD multivariate (MSL, PSM, SMAP, SMD, SWaT) SMAP MSL

Metrics: Affiliation-F1 F1_T Standard-F1 VUS-PR per-series runtime

Baselines: TimeRCD DADA TS-Pulse MOMENT TimesFM Chronos TimeMOE TranAD USAD OmniAnomaly LOF IForest

Builds on: Time-RCD DLinear MAE TSB-AD VLM4TS

Closest papers by meaning

• Towards Zero-Shot Anomaly Detection and Reasoning with Multimodal Large Language Models (CVPR 2025 · similarity 0.87)
• A Unified Reasoning Framework for Holistic Zero-Shot Video Anomaly Analysis (NeurIPS 2025 · similarity 0.86)
• UniVAD: A Training-free Unified Model for Few-shot Visual Anomaly Detection (CVPR 2025 · similarity 0.85)
• Supervision Recovery for Time Series Anomaly Detection via Counterfactual Pairing (NeurIPS 2026 · similarity 0.85)
• The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark (NeurIPS 2024 · similarity 0.85)
• Advancing Video Anomaly Detection: A Concise Review and a New Dataset (NeurIPS 2024 · similarity 0.85)