acceptodds
Under review as a conference paper at ICLR 2027

Beyond a Fixed Patch Scale: Scale-Decoupled Multi-Granularity Evidence Fusion for Time-Series Anomaly Detection

Abstract

Patch-based time-series anomaly detection typically relies on a predefined patch length, thereby restricting anomaly observation to a single temporal receptive field. However, real-world anomalies exhibit substantial heterogeneity in duration and contextual dependency, making a fixed scale prone to scale mismatch. To address this issue, we propose MGF-AD (Multi-Granularity Evidence Fusion for Scale-Uncertain Time-Series Anomaly Detection), which interprets patch length as the temporal granularity of anomaly observation. MGF-AD independently learns scale-specific normal patterns at multiple temporal scales, projects the resulting patch-level anomaly evidence onto a shared timeline via patch-to-point alignment, and performs per-scale score calibration followed by test-label-free fusion. This design explicitly decouples scale-specific normality modeling from cross-scale anomaly decision making. On TSB-AD-U and TSB-AD-M, MGF-AD achieves an overall advantage over the globally best fixed scale on the primary detection metrics and partially closes the performance gap between a globally fixed scale and the per-series optimal scale. Further analysis shows that anomaly evidence from different temporal granularities is complementary, suggesting that replacing fixed-scale selection with cross-scale evidence integration can reduce the sensitivity of patch-based TSAD to a single predefined temporal granularity.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Anomaly Detection, Multi-Granularity Evidence Fusion, Scale Mismatch, Patch-Based Representation Learning, Scale-Decoupled Normality Modeling, Multi-Scale Anomaly Detection

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_NQBGOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses scale mismatch in patch-based TSAD, where a single predefined patch length may not fit anomalies of different durations. The proposed MGF-AD works independently at each of several patch lengths. At each scale, it trains a 1D-CNN patch encoder with neighborhood-consistency metric learning and a temporal-continuity auxiliary loss, builds a bank of normal prototypes via k-means, and scores each test patch by its cosine distance to the nearest prototype. Patch scores are averaged onto timestamps, min–max normalized per scale and sequence, and fused by an equal-weight average.

Critical strengths
  • The evaluation follows a reliable protocol. The TSB-AD benchmark, separate tuning and evaluation splits, and no point adjustment.
  • Controlled comparisons are included. Fixed-scale variants and a homogeneous ensemble with matched branch counts.
Critical weaknesses
  • About motivation, short patches lack context, and long patches dilute short anomalies. However, there is no evidence that scale mismatch is a major failure source of patch-based TSAD, such as how much the performance of existing methods varies with patch length.
  • Scale mismatch is hard to regard as an unsolved problem. The actual gap would be why score-level fusion is needed instead of representation-level multi-scale modeling, and the paper does not discuss this. For the same reason, Contribution 1 (“identifying the scale-mismatch problem”) is not new.
  • The paper criticizes a fixed patch scale, yet the solution is itself a predefined, fixed set of scales {64, 96, 128}. The problem of choosing one L simply becomes the problem of choosing a set of L. Table 4 shows performance varies non-monotonically with the scale combination (on TSB-AD-U, {32, 64, 96, 128} < {64, 96, 128}).
  • The stated challenge is that the optimal scale for each series or segment is unknown at test time. The solution, however, averages all scales with equal weights for every series and timestamp. It neither estimates nor adapts to the optimal scale; it only hedges.
  • “Decoupling” scale-specific normality modeling from cross-scale decision making is a classic late-fusion ensemble of independently trained per-scale models. The paper never argues why this is better than joint multi-scale representation learning. On TSB-AD-M, the gain over a homogeneous ensemble (3×Scale-96) is small (Table 5).
  • Patch-to-point alignment and per-scale calibration are necessary preprocessing for averaging scores across scales, not designs derived from the stated challenge.
  • Min-max calibration is offline and uses statistics of the entire test sequence. So although the method is test-label-free, it still uses test-time information.
  • Each scale branch (a lightweight patch encoder with nearest normal-prototype scoring) is nearly identical to PaAno as described in the introduction. Beyond repeating it across scales and averaging, the novelty over PaAno is unclear.
Questions for authors
  • See the weaknesses.
  • Overall, the method reduces to a score-level ensemble of PaAno-like models over a hand-picked set of scales. It does not resolve the scale uncertainty that motivates it.
Reviewer_NXRROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper attempts to address the issue of scale mismatch caused by a fixed patch length in patch-based time series anomaly detection. The authors propose MGF-AD, a multi-granularity evidence fusion framework, which reduces the dependence of detection results on a single preset scale through independent modeling at multiple scales, anomaly score alignment and calibration, and cross-scale fusion without the need for test labels.

Critical strengths
  1. It identifies the scale mismatch problem where fixed patch lengths struggle to adapt to anomalies of varying durations, and sets a clear research focus.
  2. It proposes combining detection results across different scales. The approach is straightforward and improves detection performance.
  3. Control experiments provide evidence for the complementarity of anomaly information across different scales, supporting the method’s effectiveness.
Critical weaknesses
  1. The methodological novelty is limited, as MGF-AD mainly extends PaAno (ICLR 2026) to multiple patch scales and combines anomaly scores through standard normalization and averaging, as already explored in AutoTSAD (PVLDB 2024).
  2. The method offers poor cost-effectiveness. Its model size and runtime are nearly three times those of single-scale baselines, but the improvement in detection performance on TSB-AD-M is small and does not justify the extra cost.
  3. The underlying problem remains unresolved. The observation window sizes must still be manually selected in advance, and the calibration step requires the full test sequence, making the method unsuitable for direct use in real-time detection scenarios.
  4. Key design choices are largely empirical. For example, simply averaging results across different scales lacks theoretical justification, and its effectiveness beyond the TSB-AD benchmarks has not been verified.
Questions for authors
  1. What is the core novelty of MGF-AD compared with existing multi-scale methods? Does it still outperform a multi-scale version of PaAno?
  2. Can the authors provide statistical significance tests to show that the small performance gains on TSB-AD-M are reliable?
  3. Does MGF-AD still outperform single-scale methods under the same computational budget?
  4. Can MGF-AD maintain its performance when anomaly scales change or online calibration is used?
Reviewer_D65nOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The study proposes a new model for time-series anomaly detection, named MGF-AD. The main idea is to use patches at multiple temporal scales to capture different patterns of normal behavior and combine their anomaly scores to improve detection performance. This approach is motivated by the fact that anomalies can vary in duration, making different patch lengths more suitable for detecting different types of anomalies. The proposed MGF-AD model independently trains a 1D-CNN encoder and a two-layer MLP projection head for each patch scale. The training process combines two loss terms. The first encourages neighboring patches to have similar representations (triplet-based metric). The second is a temporal-continuity loss that trains a binary classifier to distinguish temporally adjacent patch pairs from non-adjacent ones. After training, only the CNN encoder is retained to extract patch representations for anomaly detection. The model first constructs a memory bank of normal patch representations for each scale using K-means clustering. Test patches are then compared with the stored normal patterns using cosine similarity to calculate anomaly scores. These patch-level scores are averaged over overlapping patches to obtain time-point anomaly scores. Finally, the scores from different scales are normalized and averaged to produce the final anomaly scores.

Critical strengths
  • (1) MGF-AD demonstrates strong anomaly detection performance on both univariate and multivariate benchmarks. I also appreciate the use of appropriate evaluation metrics and the avoidance of point adjustment.

 

  • (2) The experimental evaluation is comprehensive, covering multiple datasets and comparisons with a wide range of baseline methods. The authors also provide relevant ablation studies to assess the contributions of the proposed components.

 

  • (3) The motivation for using multiple patch scales in time-series anomaly detection is reasonable and clearly presented.
Critical weaknesses
  • (1) Several components of the proposed MGF-AD model have already been explored in the existing literature. In particular, the representation-learning procedure applied independently at each patch scale shares similarities with PaAno (ICLR 2026) [1]. Both methods employ a CNN encoder, an MLP projection head, triplet-based learning to encourage similarity between temporally neighboring patches, an auxiliary temporal-continuity objective, and memory-based anomaly scoring. The main distinction is that MGF-AD applies this procedure at multiple patch scales and combines the resulting anomaly scores. However, multi-scale modeling for time-series anomaly detection is not a new concept, as it has already been investigated in approaches such as the multi-scale recurrent autoencoder method [2], MP-MixerNet [3], and TransDe [4]. More importantly, the method in [2] already considers anomaly evidence derived from reconstruction errors at multiple temporal scales. Therefore, I encourage the authors to more clearly distinguish their methodological contributions from these existing approaches and explain what is genuinely new. I appreciate the discussion in Sections A.2 and A.3 of the appendix. However, a more thorough comparison with existing methods and a clearer discussion of the proposed method's distinct methodological contributions are needed.

 

  • (2) The proposed method relies on three predefined patch scales (64, 96, and 128). I appreciate the ablation study in Table 4 examining the selected scales. However, the proposed approach does not provide a mechanism to automatically learn or adaptively select the most appropriate patch scales for each dataset.

 

  • (3) The Affiliation F1-score (Aff-F1) should also be included, as it is considered one of the most appropriate evaluation metrics for time-series anomaly detection.

 

  • (4) Only one method from 2026 is included in the experimental comparison. I would like to see comparisons with more recent methods, particularly those published at ICLR 2026 and NeurIPS 2025, to provide a more comprehensive evaluation of the proposed approach against recent state-of-the-art methods.
Questions for authors
  • (1) The encoder architecture is only described in the appendix. I suggest that the authors include a description of the encoder architecture in the main manuscript, while providing further implementation details in the appendix if space is limited.

 

  • (2) I suggest that the authors include Figure 5, or a similar illustration, in the main manuscript, as it clearly illustrates the main motivation behind the proposed approach.

Minor comments

  • Paragraph 3.1: be the observation at timestamp t --> be the multivariate observation at timestamp t

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.