acceptodds
Under review as a conference paper at ICLR 2027

TVF-AD: Temporal-Visual Fusion for Time-Series Anomaly Detection

Abstract

Detecting abnormal temporal dependencies is fundamental to time series anomaly detection (TSAD), yet accurately modeling temporal dependencies, especially in multivariate time series, remains an open challenge. Recent approaches leverage pretrained visual models to capture temporal structures, showing promising performance by transferring structured visual semantics to sequential data and enhancing temporal structure modeling. However, such methods often rely heavily on visual priors and may fail to capture fine-grained temporal dynamics and domain-specific dependencies inherent in time series. To address this limitation, we propose TVF-AD, a novel hybrid temporal–visual representation-based framework for TSAD that integrates general visual priors with specialized temporal representations. Our framework effectively bridges cross-modal structural knowledge and sequential dynamics, enabling more accurate and robust anomaly detection, while maintaining competitive efficiency by optimizing only lightweight temporal encoding and fusion modules. Extensive experiments on benchmark datasets demonstrate that our method consistently outperforms state-of-the-art approaches.

Primary area:
learning on time series and dynamical systems
Keywords:
time series anomaly detection, representation learning, temporal-visual fusion, pretrained visual models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_3bitOpenReview ↗
Rating
3
Confidence
1
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TVF-AD for time-series anomaly detection. It combines a lightweight temporal encoder with a frozen vision model that processes time-frequency images. Anomaly scores are computed using a normal feature bank and refined through offline post-processing with multi-scale temporal cues.

Critical strengths
  1. TVF-AD achieves strong results on both univariate and multivariate TSB-AD benchmarks, with consistent gains over temporal-only and visual-only variants.

  2. Extensive ablations, including shared post-processing controls and visual-encoder comparisons, help clarify the sources of improvement.

Critical weaknesses

1.The geometry alignment objective admits a visual-only solution as the gate approaches zero. The reported gate statistics do not establish whether the temporal branch learns complementary information.

2.The contribution of visual pretraining remains unclear, as the random-versus-pretrained encoder comparison is limited to visual-only variants rather than the full fusion model.

3.Post-processing contributes substantially to performance, but matched post-processing comparisons with competing methods are reported only on TSB-AD-M. Equivalent controls on TSB-AD-U would strengthen the attribution of gains.

Questions for authors

Please address the concerns in Critical Weaknesses.

Reviewer_3g52OpenReview ↗
Rating
3(+1)
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes TVF-AD, a time-series anomaly detector that combines a temporal CNN with frozen visual features extracted from CWT images. A learned gate fuses the representations, which are trained through visual geometry alignment and scored using memory-bank retrieval and auxiliary post-processing.

Critical strengths
  1. The method achieves strong aggregate results on both univariate and multivariate TSB-AD benchmarks across five primary metrics without point adjustment.
  2. The extensive ablations examine fusion, training objectives, variable grouping, temporal conditioning, and post-processing. Applying the same post-processing to competing methods on TSB-AD-M is a useful control.
  3. The subset results and analyses of gate behavior provide useful information about performance variation and the complementary behavior of the two branches.
Critical weaknesses
  1. The methodological novelty requires further clarification, as the visual construction follows ITF-TAD and the framework combines established encoding, gating, and retrieval components. Table 20 compares random and pretrained encoders only in visual-only variants, leaving the contribution of pretrained visual knowledge to the full fused model insufficiently isolated.
  2. The geometry objective can be minimized by reproducing the visual representation, since setting the fusion gate toward the visual branch makes the two distance matrices identical. The objective ablations do not include an untrained or no-alignment control, making it unclear why the proposed training preserves complementary temporal information.
  3. Post-processing increases univariate VUS-PR from 0.5802 to 0.6721, accounting for a substantial part of the reported improvement. However, the comparisons that apply identical post-processing to competing detectors are provided only for the multivariate benchmark.
  4. Some supplementary metrics require clarification. For example, Genesis has Standard-F1 of 0.8942 in Table 23 but PA-F1 of 0.5549 in Table 24. The manuscript does not explain the thresholding protocols that could produce this ordering.
Questions for authors
  1. Can the full fused detector be evaluated with a random visual encoder and a matched non-pretrained time-frequency encoder to isolate the benefit of visual pretraining?
  2. What prevents the geometry objective from favoring the visual-copy solution, and how does performance compare with no alignment training?
  3. Can identical post-processing be applied to the strongest univariate baselines, including PaAno and ViT4TS?
  4. How are PA-F1 thresholds selected, and how are the Genesis and CreditCard differences between Standard-F1 and PA-F1 explained?
Reviewer_NZSBOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TVF-AD converts time series into CWT time-frequency images, extracts visual features with a frozen CLIP ViT, and combines them with a lightweight temporal branch via a vector gate. Anomalies are scored with a memory bank at inference, followed by post-processing. The method performs well on the univariate and multivariate TSB-AD subsets, with ablations on variable grouping and fusion. What remains unclear is what TVF-AD adds over recent visual TSAD methods, and how its post-processing relates to the TSB-AD evaluation protocol.

Critical strengths

1.The ablations on variable grouping and on several fusion strategies give useful empirical evidence about where the method's gains come from. 2.The evaluation covers both TSB-AD subsets and includes visual-related baselines (ITF-TAD-ResNet, ITF-TAD-ViT-L, ViT4TS), a reasonable starting point for comparison.

Critical weaknesses

1.The post-processing needs to be reconciled with the TSB-AD protocol. TSB-AD's evaluation is a unified setup — standardized hyperparameter tuning, z-score normalization, and VUS-PR as the primary measure — rather than a per-method score pipeline, so the paper should state explicitly whether TVF-AD's multi-step post-processing is applied under the same protocol for every method, and whether the quantile threshold in Algorithm 3 uses test-sequence statistics. The concern is not label leakage: the quantile does not use labels. It is whether the reported comparison holds under a consistent pipeline, and the effect is material — VUS-PR on TSB-AD-U rises from 0.5802 (base score) to 0.6721 (full configuration). Table 19 helpfully applies the same post-processing to PaAno and CARLA, but that covers only two methods on the multivariate benchmark; results with and without post-processing under a consistent protocol are needed for the main baselines. 2.The comparison with visual methods is incomplete. The inclusion of ITF-TAD and ViT4TS shows the issue is not a total absence of visual baselines. But PRISM and VAN-AD are neither compared nor adequately discussed, and it is unclear why the full VLM4TS pipeline is omitted. The "Visual only" ablation uses TVF-AD's own visual branch, so it does not demonstrate competitiveness against existing visual TSAD approaches. If direct comparison is not feasible, the paper should explain the differences in task, training setup, and evaluation protocol.

Questions for authors

1.What is TVF-AD's main contribution relative to PRISM's time-series-to-image approach, and is the geometry alignment loss necessary to that contribution? 2.How does the post-processing in Algorithm 3 interact with the TSB-AD protocol — specifically, is the quantile threshold derived from test-sequence statistics, and is the same post-processing applied to all baselines? 3.Could the authors report results for the main baselines both with and without the same post-processing, so the ranking can be assessed under a consistent pipeline?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.