acceptodds
Under review as a conference paper at ICLR 2027

SCAR: Condition-Aware Retrieval of Local Normal References for Multivariate Time Series Anomaly Detection

Abstract

Industrial systems frequently switch among operating regimes, and the same local pattern may be normal under one regime yet anomalous under another. The normal reference used for anomaly scoring should therefore vary with operating conditions. Most existing methods, however, still treat the normal reference as a globally uniform object rather than one determined dynamically by query-specific conditions. We propose SCAR (Slow-fast Condition-Aware Retrieval), which dynamically constructs a condition-aware local normal reference for each query and uses it as the primary basis for anomaly scoring. SCAR decomposes each time window into a slowly varying state component and a rapidly varying residual component. It first retrieves regime-consistent candidate windows through state-level retrieval and then selects locally similar normal patches through patch-level context retrieval. SCAR then compares the query against the retrieved reference to obtain a memory distance, combines it with complementary evidence including state novelty and reconstruction diagnostics, and fuses all signals into a final anomaly score via empirical CDF calibration. On five public multivariate time-series anomaly detection benchmarks, SCAR achieves the best AUROC on every dataset as well as the highest average AUROC and AP among 24 baselines. Ablation studies show that this advantage stems from the complete design chain built around condition-aware local normal references rather than from any single component in isolation. Mechanism validation on the TEP dataset, which provides ground-truth regime annotations, further confirms that the learned state space effectively recovers true operating modes and that retrieved normal references predominantly share the same regime as the query.

Primary area:
learning on time series and dynamical systems
Keywords:
Multivariate Time Series Anomaly Detection, Condition-Aware Retrieval, Hierarchical Retrieval, Time Series Decomposition, Self-Supervised Learning

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 4 reviews, above 27% of the others.
Reviewer_jiB8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes SCAR, a method for multivariate time-series anomaly detection that addresses the problem of normal behavior differing across operating regimes. The method first decomposes each time window into slowly varying state information and rapidly varying local information, then retrieves normal reference segments in two stages based on the operating regime and local context. Finally, reference distance, state novelty, and reconstruction errors at two scales are mapped onto a common scale and combined into an anomaly score. The paper compares SCAR with 24 baselines on five main datasets and uses the TEP dataset to examine the relationship between retrieved references and actual operating regimes. Overall, the method achieves competitive detection performance.

Critical strengths
  1. The research problem is practically relevant, and the motivation is clear. When an industrial system switches operating regimes, the same local variation may change from normal to anomalous. Retrieving normal segments from similar operating regimes before making local comparisons is a reasonable and intuitive design.

  2. The main detection results are competitive. In Table 1, SCAR achieves the highest AUROC on all five datasets, with average AUROC and AP values of 0.7760 and 0.3666, respectively. In Table 2, the full method also substantially outperforms the global retrieval variant, providing experimental support for combining operating regimes with local context.

  3. The experiments cover a broad range of settings and acknowledge the method’s limitations. Beyond the main benchmarks, the paper includes experiments on 530 TSB-AD time series, an analysis of TEP with ground-truth operating-regime labels, and experiments on contamination, ablations, and computational cost. The authors also acknowledge that fusion and purification do not help in every setting, allowing readers to understand the method’s practical trade-offs.

Critical weaknesses
  1. The final scoring method is inconsistent with the explanations accompanying the case-study figures. Equation (11) on page 7 and the description on page 18 state that the final score is a fixed, equally weighted average of four diagnostic scores. Page 17 also explicitly states that prediction errors are used only for auxiliary training and do not contribute to the final score. However, Figures 5–9 on pages 29–33 show time-varying “fusion weights” and explain anomaly scores using long- and short-term prediction branches. These internal inconsistencies in the reported method and results indicate that the manuscript is not ready for submission.

  2. The claim of stable calibration across operating regimes lacks independent validation. Appendix F first uses the normal set A to determine quantile thresholds and then computes regime-specific false-positive rates on that same set A. This analysis can describe differences across operating regimes within those data, but it cannot establish that false-positive rates remain controlled on new normal data. Moreover, Table 3 reports a tail error of 0.0572 at the 95% threshold, meaning that the mean absolute deviation of regime-specific false-positive rates from the target rate of 5% is 5.72 percentage points. The standard deviation of false-positive rates across operating regimes is 7.31 percentage points. These values are large, and independent testing and appropriate comparisons are needed to support the claim of stable calibration.

  3. Variability across experiments and execution configurations are insufficiently documented, making some discrepancies difficult to explain. The first three groups of sensitivity experiments in Figure 4 on page 27 are all labeled as reusing Stage-A, yet the results corresponding to the same default parameters differ: 0.805, 0.805, and 0.798 for MSL, and 0.558, 0.571, and 0.571 for SMAP. These differences exceed rounding precision. The paper does not adequately explain these discrepancies and does not report variability across repeated runs for the main results. In particular, the full method’s average AP exceeds that of kNN by only 0.0069, so the available results are insufficient to determine whether this advantage is stable.

Questions for authors

The weaknesses above raise concerns about the correctness of the reported results. The model architecture appears to be described inconsistently across the manuscript, and the authors have not released the code. Together, these issues raise doubts about the reliability of the results, including concerns that they may be inaccurate or fabricated. The authors should provide sufficient experimental evidence to address these weaknesses.

3 comments
AuthorsOpenReview ↗

Response to Reviewer jiB8 Part1

Thank you for recognizing the motivation, detection performance, and experimental coverage of our work, and for identifying inconsistencies between the method description, case interpretations, and experimental records. Your central concern is whether these elements correspond to the same reproducible system. Accordingly, we have checked the scoring implementation, case-visualization procedure, and saved results, and supplemented them with a holdout evaluation of normal data and repeated runs under a fixed protocol. Below, we explain the verified facts, supplementary results, and aspects that still require correction.

1. Inconsistencies between final scoring and the interpretation of case figures

1.1 The auxiliary prediction task and reconstruction components in the case studies

The inconsistency you identified arose from the component names and explanations in Appendix M. The two scale-specific components used during SCAR inference are masked-reconstruction errors at patch sizes 8 and 32, measuring the recovery error of observed segments after masking. Next-patch prediction serves as an auxiliary training task during representation learning; its error does not enter the final anomaly score.

We have redrawn the case figures for all five datasets, replacing the original "Short/Long prediction" labels with Masked reconstruction (p = 8) and Masked reconstruction (p = 32), and rewritten the case analyses around the temporal behavior of events, multiscale reconstruction responses, memory deviation, and state novelty. The same 15 case windows are retained. For example, PSM C1/C2 exhibit broad score responses with a prominent scale-32 reconstruction signal; the brief disturbance in C3 is accompanied by a local peak and an increase in the scale-8 reconstruction display share. The corresponding explanations now agree with the actual scoring path.

Manuscript changes

  • Appendix B.3.3: p. 19, lines 989-990.
  • Appendix M: p. 31, lines 1622-1625.
  • Figures 5-9 and the corresponding case analyses: pp. 31-35.

1.2 Fixed equal-weight fusion and time-varying display shares

Equation (11) uses fixed equal-weight fusion. Denoting the four calibrated diagnostic components by uj(t)u_j(t), the final anomaly score is:

S(t)=14∑j=14uj(t).S(t)=\frac{1}{4}\sum_{j=1}^{4}u_j(t).

Each component's coefficient remains 1/41/4, while its value varies over time. The varying proportions in Appendix M come from post-hoc visualization. The original manuscript called them "dynamic fusion weights," without clearly distinguishing display proportions from the model's scoring coefficients. We have renamed the two panels Display share (post-hoc softmax) and Allocated display score, respectively, and added their computational definitions.

Specifically, each diagnostic component and the final score separately use points labeled as normal in the test set of the same run as the percentile reference. A moving average of width 5 produces the component display curves qj(t)q_j(t) and the final display curve g(t)g(t). A softmax with temperature 0.1 is applied to the component curves, followed by a moving average of width 7 and normalization at each time step, yielding the display shares aj(t)a_j(t) in row (d). The thickness of each colored band in row (c) is:

bj(t)=aj(t)g(t),∑jbj(t)=g(t).b_j(t)=a_j(t)g(t),\qquad \sum_j b_j(t)=g(t).

Thus, the stacked panel shows a post-hoc allocation of the final display curve across diagnostic signals. The additive terms in the actual score are uj(t)/4u_j(t)/4, which have a different meaning from bj(t)b_j(t) in the figure. The percentile transformations and softmax above are used for plotting; the detector still calibrates against training references and uses the equal-weight scoring in Equation (11). The revised legends, panel names, and case explanations all use this consistent definition.

Manuscript changes

  • Appendix B.5.3: p. 20, lines 1070-1072.
  • Appendix M: p. 31, lines 1626-1632.
  • Figures 5-9, rows (b)-(d), and the corresponding analyses: pp. 31-35.
AuthorsOpenReview ↗

Response to Reviewer jiB8 Part2

2. Lack of held-out normal-data validation for cross-regime calibration

Thank you for pointing out this issue. The original Appendix F determined thresholds and calculated false-positive rates on the same normal set, without examining threshold performance on new normal data. Your interpretation of Table 3 is also correct: a target deviation of 5.72 percentage points and a cross-regime standard deviation of 7.31 percentage points are insufficient to support the claim of "stable calibration across operating regimes."

To address this issue, we added a temporal holdout evaluation of normal data covering six TEP modes: a normal segment outside the training data determines the alarm threshold, and a separate, later normal segment that has not participated in training, calibration, or threshold setting is used to evaluate the false-positive rate. Below, we refer to these two segments as the "threshold-setting set" and the "held-out evaluation set," respectively. We compare two settings that use either the normal training set or the threshold-setting set as the reference for score calibration; both determine the final alarm threshold on the threshold-setting set and evaluate it on the same held-out evaluation set.

Score calibration referenceTarget false-positive rateOverall false-positive rate on the held-out evaluation setStandard deviation of false-positive rates across six modesMean absolute target deviation
Normal training set5%0%0 percentage points5 percentage points
Normal training set1%0%0 percentage points1 percentage point
Threshold-setting set5%0%0 percentage points5 percentage points
Threshold-setting set1%0%0 percentage points1 percentage point

On the held-out normal time segments from the six modes in this evaluation, neither calibration-reference setting produced false alarms, and the observed false-positive rates were below the specified 5% and 1% levels. This adds an evaluation of the thresholds on new normal time segments; however, the target deviations remain 5 and 1 percentage points, indicating conservative thresholds and not establishing that the false-positive rates precisely match the targets or that either calibration setting is superior. We have accordingly corrected the wording "stable calibration across operating regimes," clarifying that this holdout evaluation supports the statement that "on the evaluated held-out normal time segments from the six modes, false-positive rates did not exceed the specified targets."

Manuscript changes (all marked in blue; page and line numbers refer to the revised manuscript compiled for this revision):

  • p. 9, Section 4.4, lines 466-470; the calibration interpretations and note for Table 3, lines 480-485: distinguish the original same-set statistics from the added temporal holdout evaluation, retaining the original values and correcting their meaning.
  • p. 23, Appendix F, lines 1240-1241; pp. 25-26, new Appendix F.1, lines 1326-1354 (including p. 25, Table 8, lines 1342-1349): describe the holdout evaluation procedure and report false-positive rates, cross-mode standard deviations, and target deviations.
  • p. 27, Appendix G, lines 1409-1412; p. 26, Figure 3 caption, lines 1398-1399: correct the conclusions drawn from the original tail statistics and note that the figure uses the same normal set for threshold setting and evaluation.
AuthorsOpenReview ↗

Response to Reviewer jiB8 Part3

3. Differences at sensitivity-analysis default points and AUROC/AP variation across seeds

To address the consistency of default points and AP variation across runs, we summarize AUROC/AP for five seeds on each dataset and report the means and sample standard deviations. Metrics use a 0-1 scale.

SeedMSL (AUROC / AP)PSM (AUROC / AP)SMAP (AUROC / AP)SMD (AUROC / AP)SWaT (AUROC / AP)Macro-average (AUROC / AP)
70.804949 / 0.2996550.806539 / 0.5761690.593723 / 0.1592950.824955 / 0.1942260.859023 / 0.6126540.777838 / 0.368400
170.783341 / 0.2610140.803330 / 0.6064820.642018 / 0.2056830.818582 / 0.1866980.871892 / 0.6388670.783833 / 0.379749
420.790831 / 0.2668450.812555 / 0.5933960.561231 / 0.1425090.819608 / 0.2074020.861967 / 0.7617720.769238 / 0.394385
20260.707529 / 0.2066550.790244 / 0.5741400.621037 / 0.1657160.814948 / 0.1924830.861082 / 0.7541070.758968 / 0.378620
34070.791219 / 0.2702580.803769 / 0.5925170.594681 / 0.1614500.818000 / 0.1879360.844616 / 0.7449780.770457 / 0.391428
DatasetAUROC mean ± standard deviationAP mean ± standard deviation
MSL0.775574 ± 0.0388300.260885 ± 0.033791
PSM0.803287 ± 0.0081670.588541 ± 0.013432
SMAP0.602538 ± 0.0306050.166931 ± 0.023388
SMD0.819219 ± 0.0036460.193749 ± 0.008242
SWaT0.859716 ± 0.0097940.702476 ± 0.070891
Macro-average0.772067 ± 0.0094060.382516 ± 0.010517

Sensitivity-analysis default points. With the same default hyperparameters, different random runs may still produce different scores. The maximum differences among the default points for MSL and SMAP are 0.6846 and 1.2855 percentage points, respectively; the corresponding AUROC standard deviations in the five-seed results above are 3.8830 and 3.0605 percentage points, respectively. Numerically, both default-point differences are smaller than the corresponding cross-seed standard deviations in the table.

Whether the AP advantage disappears with variation across runs. Your concern is whether the macro-average AP gain of only 0.69 percentage points in the original manuscript persists across runs. Taking the original kNN result of 35.97% as the reference, the macro-average APs of the five seeds above are 36.84%, 37.97%, 39.44%, 37.86%, and 39.14%, respectively. All exceed this reference, with gains ranging from 0.87 to 3.47 percentage points. The five-seed macro-average AP is 38.25% ± 1.05 percentage points, with an average gain of 2.28 percentage points. Thus, across these five sets of results in the table, the macro-average AP advantage over the original kNN result does not reverse.

4. Reliability of results and reproducibility materials

We submitted the code as supplementary material with the original submission. To make it easier for you to inspect the code, we have further organized it and provided the following anonymous access link to the GitHub repository, which can be browsed and downloaded without logging in:

Anonymous code repository: https://anonymous.4open.science/r/scar-review-code-0E9B/

The repository includes the model, retrieval and scoring implementations, configurations for the five main datasets, training and evaluation entry points, environment instructions, and calibration tests, supplementing the materials submitted with the paper.

We plan to release the model weights on Hugging Face in the future. At present, to comply with anonymous submission requirements and avoid inadvertently revealing author identities through hosting accounts or file metadata, we are not providing a weight-download link; we will add it after completing the anonymity checks. We apologize for any inconvenience and appreciate your understanding. The current anonymous code repository does not contain model weights.

Reviewer_r3jAOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SCAR, a multivariate time-series anomaly detection framework that constructs query-specific normal references conditioned on operating state and local temporal context. It decomposes input windows into slow state and fast residual components, uses state representations to filter candidate reference windows, and retrieves context-matched patches within those candidates. Anomaly scores combine content-space distances to retrieved references, state novelty, and multi-scale reconstruction errors through empirical-CDF calibration and averaging.

The main contribution is the integration of slow–fast representation learning with hierarchical, condition-aware reference retrieval. Evaluation against 24 baselines on five multivariate benchmarks reports the best AUROC on each dataset and the highest average AUROC and AP. Validation additionally includes component ablations, a six-mode TEP study examining regime alignment and retrieval behavior, evaluation on 530 TSB-AD series, and analyses of memory robustness, sensitivity, and computational cost.

Critical strengths
  1. Well-motivated and coherent method. SCAR addresses the ambiguity of local patterns across operating regimes through a clear hierarchy of state filtering, context matching, and content-distance scoring. The design connects naturally to the problem, and the retrieval ablations support the usefulness of combining state and context information.

  2. Strong overall empirical performance. SCAR achieves the highest AUROC on all five main benchmarks against 24 baselines. Evaluation on 530 additional TSB-AD series broadens the evidence for its applicability, although AP gains are less consistent.

  3. Useful analysis beyond aggregate metrics. The six-mode TEP study provides direct evidence that the learned state space and retrieved references reflect operating conditions. Component ablations, diagnostic decomposition, and memory-robustness analyses further help explain the method’s behavior and practical tradeoffs.

Critical weaknesses
  1. The incremental contribution of memory retrieval remains insufficiently isolated. The ablations support hierarchical retrieval over global and single-level alternatives. However, given the strong standalone performance of state novelty and reconstruction diagnostics, a matched comparison against their fusion without the memory branch would clarify how much condition-aware retrieval adds to the overall detector.

  2. The claim of stable cross-regime calibration needs qualification. The reported mode-wise false-positive rates show appreciable variation across operating conditions. Moreover, empirical CDFs are fitted on memory-construction data without self-exclusion. This does not establish that calibration is invalid, but evaluation on held-out normal samples would strengthen the claim.

  3. The TEP analysis does not fully connect regime alignment to detection improvements. It provides useful evidence that retrieved references generally match the query’s operating mode. Comparing fault–normal separation across retrieval strategies within the same TEP setup would more directly test whether this alignment reduces misleading matches, complementing the detection ablations already provided on the main benchmarks.

Questions for authors

see weaknesses. If all of my concerns are solved, I will consider raising my rating.

5 comments
AuthorsOpenReview ↗

Response to Reviewer r3jA Part1

Thank you for recognizing the motivation, condition-aware retrieval design, and mechanism analysis of our work, and for pointing out that the incremental contribution of retrieval, cross-regime calibration, and evidence of fault discrimination on TEP require further clarification. We address these three issues below.

1. The Incremental Contribution of Retrieval Beyond State Novelty and Reconstruction

The comparison you suggested directly examines whether the retrieval-derived memory deviation provides additional anomaly evidence beyond state novelty and reconstruction. Comparisons among global, state-only, and context-only retrieval evaluate how normal references are selected, whereas the present experiment isolates the contribution of memory deviation to the final anomaly score. Specifically, we remove the memory-deviation component from score fusion while retaining state novelty and reconstruction, allowing us to assess its incremental contribution to the complete detector.

We compare the following two scoring combinations:

  • Full SCAR: an equally weighted average of four calibrated components: memory deviation, state novelty, reconstruction error at scale 8, and reconstruction error at scale 32.
  • State novelty + reconstruction: memory deviation is removed, and the remaining three components are averaged.
DatasetFull SCAR, seed 17 (AUROC / AP, %)State novelty + reconstruction (AUROC / AP, %)AUROC difference (percentage points)AP difference (percentage points)
MSL78.33 / 26.1077.83 / 25.23+0.50+0.87
PSM80.33 / 60.6580.43 / 56.70-0.10+3.95
SMAP64.20 / 20.5760.83 / 16.29+3.37+4.27
SMD81.86 / 18.6780.90 / 17.59+0.96+1.08
SWaT87.19 / 63.8984.15 / 54.62+3.04+9.27
Macro-average78.38 / 37.9776.83 / 34.09+1.55+3.89

In this comparison, full SCAR improves macro-average AUROC and AP by 1.55 and 3.89 percentage points, respectively. It achieves higher AUROC on four of the five datasets and higher AP on all five, with particularly notable improvements on SMAP and SWaT. These results indicate that retrieval-derived memory deviation provides complementary anomaly evidence beyond state novelty and reconstruction, although its contribution varies across datasets. The effectiveness of the two-level retrieval strategy itself is separately evaluated through the retrieval ablations in Table 2 of the main paper. Figures 5–9 in Appendix M (Additional qualitative case studies) further illustrate the complementary roles of these diagnostic signals. Across different datasets and different event intervals within the same dataset, memory deviation, state novelty, and multiscale reconstruction exhibit different response strengths and combinations. Reconstruction or state novelty is more prominent in some events, whereas memory deviation is more pronounced in others, providing qualitative evidence of the potential complementarity of these diagnostic signals. The contribution of memory deviation should therefore be understood in the context of specific events, rather than requiring it to dominate at every time interval.

AuthorsOpenReview ↗

Response to Reviewer r3jA Part2

2. Cross-Regime Calibration Requires Evaluation on Held-Out Normal Data

Your concern is valid: fitting empirical CDFs on the normal data used to construct the memory bank and then evaluating false-positive rates on those same data does not test performance on new normal data. In the original Table 3, at a target false-positive rate of 5%, the mean absolute deviation from the target is 5.72 percentage points, and the standard deviation of false-positive rates across regimes is 7.31 percentage points. These values are insufficient to support the original claim of "stable calibration across operating regimes."

We added an evaluation on held-out normal time segments from the six-regime TEP dataset. The normal training segment is used for model training and memory construction; a separate normal segment outside training is used to determine alarm thresholds and is referred to as the "threshold-setting set"; another, chronologically later normal segment is used solely to evaluate false-positive rates and is referred to as the "held-out evaluation set." Neither the threshold-setting set nor the held-out evaluation set enters the memory bank. A gap of 127 points is left between segments to prevent windows of length 128 from overlapping across segments. Consequently, queries from these two segments cannot retrieve their own windows from the memory bank.

We compared two settings that use either the normal training set or the threshold-setting set as the reference for score calibration. Both settings determine the final alarm thresholds on the threshold-setting set and are then evaluated on the same held-out evaluation set.

Score calibration referenceTarget false-positive rateOverall false-positive rate on the held-out evaluation setStandard deviation of false-positive rates across six regimesMean absolute deviation from the target
Normal training set5%0%0 percentage points5 percentage points
Normal training set1%0%0 percentage points1 percentage point
Threshold-setting set5%0%0 percentage points5 percentage points
Threshold-setting set1%0%0 percentage points1 percentage point

Neither setting produced any above-threshold alarms on the held-out normal time segments from the six regimes in this evaluation, and the observed false-positive rates were below the specified targets of 5% and 1%. The deviations from the targets remain 5 and 1 percentage points, indicating conservative thresholds rather than an exact match to the target false-positive rates. We have revised the corresponding conclusion to "the false-positive rates did not exceed the specified targets on the evaluated held-out normal time segments from the six regimes," aligning the statement with the scope of the actual validation.

Revision locations (in the same revised manuscript referenced in our response to jiB8; the corresponding changes are marked in blue):

  • Page 9, Section 4.4, lines 466–470; the calibration interpretation and note for Table 3, lines 480–485: distinguish the original statistics computed on the same data from the new temporal holdout evaluation and correct the interpretation of the original statistics.
  • Page 23, Appendix F, lines 1240–1241; pages 25–26, Appendix F.1, lines 1326–1354 (including Table 8 on page 25, lines 1342–1349): describe the holdout evaluation procedure and report false-positive rates, their standard deviation across regimes, and deviations from the targets.
  • Page 27, Appendix G, lines 1409–1412; page 26, the caption of Figure 3, lines 1398–1399: correct the interpretation of the original tail statistics and clarify that the original figure uses the same normal set for threshold determination and evaluation.
AuthorsOpenReview ↗

Response to Reviewer r3jA Part3

3. Comparing Fault–Normal Separation Under the Same TEP Setup

The question you wish to examine is whether retrieving references from the same operating regime further helps distinguish faulty from normal data. Merely counting whether the regime labels of retrieved references match does not directly answer this question.

We added a comparison of four retrieval strategies under the same TEP setup: global retrieval, state-only filtering, context-only matching, and SCAR's two-level retrieval. All four strategies share the same frozen encoder with seed=42, a mixed normal memory bank containing six regimes, and the same held-out normal time segments and 168 fault files (six regimes, with 28 fault files per regime).

Due to the labeling limitations of the TEP dataset used in this study, we have regime and fault-file identifiers but lack reliable normal/anomalous labels for pointwise evaluation, preventing direct computation of pointwise AUROC/AP. We therefore designed a proxy metric, the "fault–normal separation gap," to examine whether different retrieval strategies produce a more pronounced elevation in scores for fault files relative to normal data from the same regime. For each fault file, we compute the mean of its memory deviation component and subtract the corresponding mean on the held-out normal segment from the same regime:

=\operatorname{mean}\bigl(d_{\mathrm{memory}}(X_{m,f})\bigr) -\operatorname{mean}\bigl(d_{\mathrm{memory}}(X_{m,\mathrm{normal}})\bigr).$$ Here, $m$ denotes the regime, $f$ denotes the fault file, and $d_{\mathrm{memory}}$ is the raw memory deviation score. Put simply, we compare "how much higher the mean score of a fault file is than that of normal data from the same regime": a larger gap means that their mean scores are further apart. We first average over the 28 fault files within each regime and then take an equally weighted average across the six regimes to obtain the results below. Values such as 645.49 and 708.97 in the table are differences in raw scores, not percentages or detection accuracies. | Retrieval method | Mean score gap between faulty and normal data | Relative increase in the separation gap of SCAR over this method | Proportion of references retrieved for fault queries that belong to the same regime | | --- | ---: | ---: | ---: | | Global retrieval | 645.49 | 9.84% | 89.09% | | State-only filtering | 706.66 | 0.33% | 84.22% | | Context-only matching | 652.69 | 8.62% | 89.54% | | SCAR two-level retrieval | 708.97 | — | 84.40% | The results show that, with the same model, memory bank, and evaluation data, SCAR produces a larger mean score gap between faulty and normal data: the gap increases by 9.84% over global retrieval, with increases in the average results for all six regimes and in 145 of the 168 fault files. State-only filtering already achieves results close to full SCAR, indicating that state filtering plays the main role in this TEP experiment, while context matching provides a small additional benefit. Meanwhile, the proportions of same-regime references for global retrieval and context-only matching are 89.09% and 89.54%, respectively, higher than SCAR's 84.40%, yet their fault–normal score gaps are smaller. This shows that retrieving references from the same regime does not automatically imply better fault discrimination; the new separation metric provides evidence that regime-matching rates alone cannot offer. Thus, in response to your suggestion to compare different retrieval strategies under the same TEP setup, this experiment shows that condition-aware retrieval increases fault–normal separation on this proxy metric, providing direct comparative evidence for the link between the retrieval mechanism and gains in fault discrimination. Revision locations (all additions are marked in blue): - Page 26, Appendix F.2, lines 1356–1372: the shared TEP setup, labeling limitations, and definition of the fault–normal separation proxy in Eq. (36). - Page 26, Table 9, lines 1374–1382: results for the four retrieval strategies, including separation gaps, SCAR's relative increases, and same-mode reference rates. - Page 26, Appendix F.2, lines 1385–1398: definitions of the reported percentages and analysis linking retrieval to fault–normal separation.
Reviewer_VhrYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

In industrial systems, the same local waveform may be normal in one operating regime but anomalous in another. Therefore, selecting normal references across operating regimes may interfere with anomaly detection. To address this, this paper proposes SCAR (Slow-fast Condition-Aware Retrieval), which constructs normal references based on the query's operating regime and local temporal context. This method first decomposes each window into a slowly varying state and a rapidly varying residual. It uses the state descriptor to select windows from similar operating regimes from the normal reference memory bank. Then, within these windows, it selects reference patches based on local context similarity. Finally, the memory deviation, state novelty, and the reconstruction errors are mapped to a common scale using empirical CDFs, and then combined using an unweighted mean to obtain the anomaly score. Experimental results show that on five multivariate time series datasets, SCAR achieves the highest AUROC on each dataset and the highest macro-averaged AUROC and AP compared with 24 baselines. On the TEP dataset with ground-truth regime annotations, the experiments demonstrate that the learned state representation can distinguish operating regimes and that the retrieved references are mostly from the same regime as the query.

Critical strengths
  1. The experiments are comprehensive. Table 1 compares 24 baselines, including classical shallow detectors, reconstruction- and forecasting-based models, Transformer and general time-series backbones, decomposition and time-frequency methods, and memory-, prototype-, and retrieval-based methods. Five datasets cover spacecraft telemetry, server monitoring, and industrial control, demonstrating the effectiveness of the method on different domains.
  2. The mechanism validation is targeted. On the TEP dataset with ground-truth regime annotations, the state space learned by the method is verified to be able to distinguish different operating modes. Ablation studies also support the role of each component.
Critical weaknesses
  1. The related work is insufficient. The first claimed contribution is the argument that the normal reference for anomaly scoring should be condition-aware. However, this point has already been studied in prior work [1][2][3][4]. The paper does not cite these papers, discuss them, and compare against them.
  2. There are inconsistencies in the paper. For instance, Appendix B.3.3 clearly states that next-patch prediction is an auxiliary task during the training phase, and its error does not contribute to the final score. However, Appendix M repeatedly uses short-/long-horizon prediction branches to explain the detection results. Additionally, the fusion defined in Eq. 11 and Appendix B.5.3 is an unweighted mean, but Appendix M mentions a dynamic fusion weight. These inconsistencies affect the understanding of the method and the case analysis.
  3. The advantage in AUROC is not equally reflected in AP. In Table 1, SCAR ranks first in AUROC on all 5 datasets, but its AP is the best only on MSL. Although the macro-averaged AP is the highest, it is only 0.007 higher than the second-best kNN (0.3666 vs 0.3597). However according to Table 18, the inference time per point of SCAR is about ten times longer than that of PaAno, PUAD, and PGRF-Net.

[1] Song X, Wu M, Jermaine C, et al. Conditional anomaly detection[J]. IEEE Transactions on knowledge and Data Engineering, 2007, 19(5): 631-645.

[2] He Q P, Wang J. Fault detection using the k-nearest neighbor rule for semiconductor manufacturing processes[J]. IEEE transactions on semiconductor manufacturing, 2007, 20(4): 345-354.

[3] Ma H, Hu Y, Shi H. A novel local neighborhood standardization strategy and its application in fault detection of multimode processes[J]. Chemometrics and Intelligent Laboratory Systems, 2012, 118: 287-300.

[4] Chen Z, Liu C, Ding S X, et al. A just-in-time-learning-aided canonical correlation analysis method for multimode process monitoring and fault detection[J]. IEEE Transactions on Industrial Electronics, 2020, 68(6): 5259-5270.

Questions for authors
  1. Could you cite the references listed in weakness 1 and discuss them? Does the claim of the first contribution need to be revised accordingly?
  2. Could you explain the inconsistency mentioned in weakness 2? Should the figures and case analysis in Appendix M be regenerated using the current method?
  3. In Table 1, how much of the 22 percent improvement over PaAno comes from the idea of filtering candidates by operating regime? Could you conduct an experiment that adds a simple state filter to PaAno?
5 comments
AuthorsOpenReview ↗

Response to Reviewer VhrY Part1

Thank you for recognizing the breadth of our experiments and the mechanism validation, and for your suggestions on related work, case interpretation, and the performance-cost trade-off. We have added the relevant references, revised our first contribution and case analyses, and conducted a simple state-filtering comparison for PaAno.

Revisions are marked in blue. All page and line numbers below refer to the revised manuscript.


1. Related work and contribution

Regarding Question 1 and Critical Weakness 1.

Added references. We have added and discussed all four references you identified:

  • Song et al. (2007): modeling indicator attributes conditioned on environmental attributes.
  • He and Wang (2007): applying the k-nearest neighbor rule to fault detection in semiconductor manufacturing.
  • Ma et al. (2012): local neighborhood standardization for multimode processes.
  • Chen et al. (2020): combining just-in-time learning with canonical correlation analysis to construct local monitoring models using relevant historical samples.

These studies establish the value of conditioning and local modeling under heterogeneous operating conditions.

Revised contribution. We have revised our first contribution to focus on jointly considering operating-state compatibility and local temporal-context relevance when selecting a normal reference for each query:

We highlight the importance of jointly considering operating-state compatibility and local temporal-context relevance when constructing normal references for industrial multivariate time-series anomaly detection. These two criteria guide the selection of a query-specific normal reference, helping reduce misleading anomaly assessments caused by cross-regime or context-mismatched normal patterns.

The revised Related Work also acknowledges that existing neighbor- and memory-based methods can select different reference instances for different queries. SCAR's specific design applies the two criteria in successive retrieval steps: it first filters candidate windows by operating-state compatibility, then retrieves patch-level references by local temporal-context relevance. The discussion now focuses on the criteria and procedure for reference selection.

Manuscript changes

  • Introduction, contribution 1: p. 3, lines 108-112.
  • Related Work: pp. 3-4, lines 147-156, 175-182, and 193-198.
  • Four added references: pp. 11-13.

2. Prediction and reconstruction

Regarding Question 2: the scoring path in the case studies.

Scoring path. The inconsistency you identified arose from the component names and explanations in Appendix M. The two scale-specific diagnostics used during SCAR inference are masked-reconstruction errors at patch sizes 8 and 32. They measure the recovery error of observed segments after masking. Next-patch prediction is an auxiliary representation-learning objective during training; its error does not enter the final anomaly score.

We have redrawn the case figures for all five datasets, replacing "Short/Long prediction" with Masked reconstruction (p = 8) and Masked reconstruction (p = 32). We have also revised the analyses around event timing, multiscale reconstruction responses, memory deviation, and state novelty, retaining the same 15 case windows. For example, PSM C1/C2 exhibit broad score responses with a prominent scale-32 reconstruction signal, whereas the brief disturbance in C3 is accompanied by a local peak and an increased scale-8 reconstruction display share. The explanations now correspond to the actual scoring path.

Manuscript changes

  • Appendix B.3.3: p. 19, lines 989-990.
  • Appendix M: p. 31, lines 1622-1625.
  • Figures 5-9 and their case analyses: pp. 31-35.

3. Equal-weight fusion and display shares

Regarding Question 2: the interpretation of the visualization panels.

Actual scoring. Equation (11) uses fixed equal-weight fusion. With four calibrated diagnostics denoted by uj(t)u_j(t), the final anomaly score is

S(t)=14∑j=14uj(t).S(t)=\frac{1}{4}\sum_{j=1}^{4}u_j(t).

Each coefficient remains 1/41/4, while the diagnostic values vary over time. The varying proportions in Appendix M come from post-hoc visualization. Calling them "dynamic fusion weights" did not clearly distinguish display proportions from scoring coefficients. We have renamed the two panels Display share (post-hoc softmax) and Allocated display score, and added their definitions.

Visualization procedure. The display panels are constructed as follows:

Display curves: Each diagnostic and the final score are separately converted to percentiles using labeled-normal test points from the same run as the reference. A moving average of width 5 produces diagnostic display curves qj(t)q_j(t) and the final display curve g(t)g(t).

AuthorsOpenReview ↗

Response to Reviewer VhrY Part2

3. Continued

Display shares, row (d): A softmax with temperature 0.1 is applied across the diagnostic curves, followed by a moving average of width 7 and normalization at each time step, yielding aj(t)a_j(t).

Allocated display score, row (c): The colored band thicknesses are defined by

bj(t)=aj(t)g(t),∑jbj(t)=g(t).b_j(t)=a_j(t)g(t),\qquad \sum_j b_j(t)=g(t).

Thus, the stacked panel shows a post-hoc allocation of the final display curve across diagnostic signals. The additive terms in the actual score are uj(t)/4u_j(t)/4, which have a different meaning from bj(t)b_j(t). The percentile transformations and softmax above are used for visualization; the detector retains calibration based on training references and the equal-weight scoring in Eq. (11). The revised legends, panel names, and case explanations consistently use this distinction.

Manuscript changes

  • Appendix B.5.3: p. 20, lines 1070-1072.
  • Appendix M: p. 31, lines 1626-1632.
  • Figures 5-9, rows (b)-(d), and accompanying analyses: pp. 31-35.

4. AP gains and inference cost

Regarding Critical Weakness 3.

Detection performance. You identify an important difference between the metrics. Table 1 reports a macro-average AP of 0.3666 for SCAR and 0.3597 for kNN, an absolute gain of 0.0069, or 0.69 percentage points. AP gains also vary across datasets. AUROC measures the overall ranking of anomalous points relative to normal points, whereas AP reflects precision across recall levels. Improvements in overall ranking do not always coincide with improvements in precision among high-scoring points. We therefore report both metrics to characterize these different aspects of detection performance.

Computational cost. In Table 18, SCAR requires 0.4630 ms/point, approximately 12.7, 9.8, and 9.1 times the inference time of PaAno, PUAD, and PGRF-Net, respectively. SCAR performs state-based candidate filtering, local-context retrieval, and multiple diagnostic computations during inference. These operations implement condition-aware reference construction and introduce additional computational cost.

Target setting. SCAR primarily targets industrial multivariate time series with varying operating regimes and complex local temporal contexts. In these settings, the same local waveform may have different normality meanings across regimes. Even within similar regimes, local dynamics can affect whether a historical segment is an appropriate normal reference for the current query. SCAR therefore first selects regime-compatible candidate windows and then retrieves context-relevant reference segments, aiming to reduce interference from cross-regime or context-mismatched normal patterns.

The additional retrieval cost serves this specific need; its value should be assessed alongside regime variability, contextual complexity, and detection gains. For tasks with stable regimes, simpler local dynamics, or strict computational budgets, lightweight methods also merit consideration.


5. Simple state filtering for PaAno

Regarding Question 3.

Experimental setup. Following your suggestion, we compared PaAno and PaAno+StatFilter on all five main datasets:

  • Shared model and memory: The same PaAno encoder trained with seed 42 and the same memory bank for each dataset.
  • State descriptors: Per-channel means, standard deviations, and trend slopes within windows of length 128, standardized using training-data parameters.
  • Filtering and scoring: Select the 50 memory entries with the closest statistical state descriptors, then apply PaAno's top-3 content-distance scoring.
  • Evaluation: The same point-level evaluation procedure for both variants.

Results. Each entry is AUROC / AP (%).

DatasetPaAnoPaAno+StatFilter
MSL71.85 / 20.2571.33 / 20.61
PSM52.24 / 34.0253.24 / 34.74
SMAP47.06 / 13.5453.50 / 16.80
SMD76.54 / 17.6078.44 / 22.50
SWaT30.67 / 9.5225.15 / 9.41
Macro-average55.67 / 18.9956.33 / 20.81

Observed gains. Statistical state filtering increases macro-average AUROC by approximately 0.66 percentage points and AP by 1.83 percentage points, calculated from unrounded values. Both metrics improve on PSM, SMAP, and SMD. On MSL, AP increases slightly while AUROC decreases slightly; both metrics decrease on SWaT. The benefits of this statistical filtering rule therefore depend on the dataset.

This experiment measures the incremental effect of adding statistical state filtering to PaAno. The approximately 22-percentage-point macro-average AUROC gap between SCAR and PaAno reported in the paper also involves differences in representation learning, local-context retrieval, and multi-component scoring. The modest macro-average gain from this simple filter does not by itself account for that overall gap; the two differences should not be interpreted as an additive decomposition of component contributions.

Thank you for the response and the experiments. Weakness 2 has been resolved. However, the following issues still exist:

  1. The revised first contribution is still not very convincing. [3][4] already select similar local normal samples for each query, while SCAR only separates operating-state compatibility and local temporal-context relevance as two successive criteria. This belongs to the method design in the second contribution.

  2. After adding state filtering to PaAno, the macro-average AUROC only has a slight improvement of 0.0066. However, in SCAR, state filtering brings an improvement of 0.1128. Both use state filtering. Why is the difference so large?

  3. The reply states that about ten times the inference time is for industrial multivariate time series with varying operating regimes. Though TEP is this kind of data, only the results of SCAR itself are reported on TEP. Could you set the thresholds as in Appendix F.1 and compare the detection rates of SCAR and the baselines on the fault sequences?

Reviewer_56pjOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper prposes a framework, called SCAR, for multivariate time-series anomaly detection that distinguishes anomalous behaviour from normal variations across different operating conditions. The key idea is to construct a query-specific normal reference rather than replying on a globally shared normal reference. The authors evaluate SCAR against 24 baselines on five multivariate time-series benchmarks.

Critical strengths
  • The paper addresses an important challenge in multivariate time-series anomaly detection: normal behaviour can vary substaintially across operating conditions.
  • The proposed framework is technically coherent in general.
  • The authors conduct experiments agains a range of anomaly detection approaches, and also carry out ablation studies for mechanism-level validation.
Critical weaknesses
  • Although the experiments compare against a range of anomaly detection methods, some very relevant existing methods are not included, such as "SARAD: Spatial Association-Aware Anomaly Detection and Diagnosis for Multivariate Time Series" in NeurIPS2024. Moreover, the experiments do not establish whether the learned representations and associated architectural complexity are necessary. For example, A simpler approach could identify operating conditions using statistical features or clustering and then perform nearest-neighbour retrieval.

  • There are inconsistent improvements across metrics and datasets. Although SCAR achieves the highest AUROC on all five datasets, it achieves the highest average precision (AP) on only one. The improvements from condition-aware retrieval are also uneven. Compared with global retrieval, the full method improves AUROC by small percentage points on PSM and SMD, but by a much larger percentage point on SWaT. This may suggest the benefits may depend on dataset characteristics.

Questions for authors
  • How does SCAR compare with a simpler condition-aware retrieval baseline that identifies operating conditions using statistical features or clustering? Moreover, can you compare with SARAD or at least discuss why the anomaly detection approach proposed in SARAD may not tackle the challenges as well as SCAR?

  • Could the authors explain why the proposed components improve AUROC without consistently improving AP? What accounts for the relatively low AP on SMD?

  • Could the authors investigate whether the different level of improvements between PSM/SMD and SWaT are related to the degree of operating-regime heterogeneity?

  • Could the authors report the mean and standard deviation of the main results across multiple random seeds?

  • Given the 9-13 times higher inference cost than retrieval-oriented baselines, could the authors clarify how retrieval time scales with memory size and whether the additional computational cost is justified by the detection gains?

3 comments
AuthorsOpenReview ↗

Response to Reviewer 56pj Part1

Thank you for recognizing the problem motivation, framework design, and experimental coverage, and for your suggestions on simple conditional retrieval, differences between metrics, operating-regime heterogeneity, training stability, and computational cost. We respond to each point below.


1. Simple conditional retrieval and SARAD

The core difference between SARAD and SCAR lies in how they model normality: SARAD detects anomalies through spatial associations among variables and their reconstruction errors, whereas SCAR retrieves suitable historical normal references for the current query based on its operating state and local temporal context. Thus, SARAD is not the method most directly related to SCAR's conditional retrieval mechanism, but it also addresses multivariate time-series anomaly detection and is a valuable comparison. We have therefore added a reproduction of SARAD.

We used the official SAR76 network, a window length of 100, seed=42, and the final weights after three training epochs. We used the paper's corresponding configurations for PSM, SMD, and SWaT, and its standard configuration for MSL and SMAP. Data were obtained through our main-experiment data loader, with the training data split chronologically into 80% for training and 20% for validation; component standardization used validation-set statistics. The following table reports point-level AUROC / AP (%) under this data-processing pipeline. These differ from the range-based metrics in the original SARAD paper, and adaptations were also made to SMD sequence organization and SWaT standardization.

DatasetSARAD reproduction (AUROC / AP)SCAR in the manuscript (AUROC / AP)
MSL53.10 / 15.7479.10 / 26.90
PSM66.73 / 43.1580.15 / 51.86
SMAP42.63 / 12.6062.74 / 16.72
SMD81.04 / 16.1981.55 / 18.23
SWaT85.51 / 66.0584.44 / 69.58
Macro-average65.80 / 30.7577.60 / 36.66

Overall, in this comparison, SCAR's manuscript results exceed our SARAD reproduction in AP on all five datasets and in AUROC on four of them. SCAR's macro-average AUROC/AP is 77.60%/36.66%, compared with SARAD's 65.80%/30.75%. The advantage is reflected in both overall ranking and precision metrics, with SWaT AUROC being the metric on which SARAD performs better.

To address your suggestion of simple statistical-feature retrieval, we use SCAR-StatFilter as a comparison: it replaces the learned state descriptors in first-stage candidate filtering with per-channel window means, standard deviations, and trend slopes, while retaining SCAR's encoder, memory entries, and subsequent context-retrieval procedure. This design directly addresses your suggestion by examining whether simple statistical state descriptions can replace learned state representations for candidate filtering.

The following table presents SCAR-StatFilter alongside SCAR's manuscript results.

DatasetSCAR-StatFilter (AUROC / AP)SCAR in the manuscript (AUROC / AP)
MSL74.81 / 25.3679.10 / 26.90
PSM68.22 / 40.6980.15 / 51.86
SMAP56.20 / 14.7162.74 / 16.72
SMD64.56 / 7.8681.55 / 18.23
SWaT80.91 / 41.7184.44 / 69.58
Macro-average68.94 / 26.0677.60 / 36.66

In these results, SCAR's manuscript AUROC and AP exceed those of simple statistical-feature retrieval on all five datasets, with macro-averages of 77.60% and 36.66%, respectively, compared with 68.94% and 26.06% for SCAR-StatFilter.

AuthorsOpenReview ↗

Response to Reviewer 56pj Part2


2. Differences between AUROC and AP, and the lower AP on SMD

AUROC and AP differ in their sensitivity to changes in ranking. AUROC measures the overall probability that an anomalous point ranks above a normal point; AP summarizes precision across recall levels and is more sensitive to normal points ranked ahead of anomalous points. A component can improve the relative ordering of many normal and anomalous points, thereby increasing AUROC. However, if some normal points remain in the high-score region or move even higher in the ranking, AP may not improve accordingly.

SCAR's components focus on operating-state representation, local-context matching, and multiscale anomaly characterization, aiming to construct more suitable normal references for queries under varying operating regimes and complex contexts. These improvements help distinguish anomalous from normal patterns, but better overall discrimination does not necessarily mean more accurate prioritized alerts. Consequently, improvements in overall ranking from these components can coexist with precision losses over some recall intervals, resulting in higher AUROC without consistent AP gains.

We have added SCAR's PR curve to Appendix N, Figure 10 (p. 37) of the revised manuscript. The gray dotted line indicates the anomaly-prevalence reference level.

Anomalies account for only approximately 4.16% of SMD, with normal points outnumbering anomalous points by approximately 23 to one. Under this class distribution, even a small proportion of high-scoring normal points can generate a substantial number of false alarms, reducing precision and AP. In the figure, full SCAR achieves an AUROC of 82.50% and an AP of 19.42%; at 50% recall, precision is 13.35%, indicating that many normal points are still included among the alerts at this operating point. The low anomaly prevalence amplifies the impact of these ranking errors on precision. Therefore, improving AP on SMD requires reducing the overlap between high-scoring normal and anomalous points, rather than improving overall ranking alone.


3. Gains across datasets and operating-regime heterogeneity

We believe that dependence on operating conditions and context is key to understanding the differences in gains: does the same local pattern require different normal references under different operating conditions?

Measurements such as flow and water level in SWaT need to be interpreted together with pump and valve states and upstream and downstream processes. For example, low flow may be normal when a flow path is closed but anomalous when water supply is expected. Matching local shapes alone may retrieve similar segments from a different operating state. SCAR first filters regime-compatible windows and then retrieves context-relevant local segments specifically to reduce this reference mismatch. This provides a plausible mechanistic explanation for the larger gains on SWaT in the original manuscript.

For PSM and SMD, one possible explanation for the smaller gains is that global retrieval already finds normal segments compatible with the current operating state and local context, leaving less room for additional conditional filtering to improve performance. SCAR focuses on settings where changes in operating state and complex temporal contexts alter the expected normal behavior: historical segments with similar shapes may have different normality meanings and need to match both operating conditions and context to serve as valid references. When this conditional dependence makes global retrieval prone to including unsuitable segments, two-stage conditional retrieval has greater scope to help.


4. Means and standard deviations across random seeds

Below, we report AUROC/AP for five seeds on each dataset and summarize their means and sample standard deviations.

SeedMSL (AUROC / AP)PSM (AUROC / AP)SMAP (AUROC / AP)SMD (AUROC / AP)SWaT (AUROC / AP)Macro-average (AUROC / AP)
70.804949 / 0.2996550.806539 / 0.5761690.593723 / 0.1592950.824955 / 0.1942260.859023 / 0.6126540.777838 / 0.368400
170.783341 / 0.2610140.803330 / 0.6064820.642018 / 0.2056830.818582 / 0.1866980.871892 / 0.6388670.783833 / 0.379749
420.790831 / 0.2668450.812555 / 0.5933960.561231 / 0.1425090.819608 / 0.2074020.861967 / 0.7617720.769238 / 0.394385
20260.707529 / 0.2066550.790244 / 0.5741400.621037 / 0.1657160.814948 / 0.1924830.861082 / 0.7541070.758968 / 0.378620
34070.791219 / 0.2702580.803769 / 0.5925170.594681 / 0.1614500.818000 / 0.1879360.844616 / 0.7449780.770457 / 0.391428
AuthorsOpenReview ↗

Response to Reviewer 56pj Part3

DatasetAUROC mean ± standard deviationAP mean ± standard deviation
MSL0.775574 ± 0.0388300.260885 ± 0.033791
PSM0.803287 ± 0.0081670.588541 ± 0.013432
SMAP0.602538 ± 0.0306050.166931 ± 0.023388
SMD0.819219 ± 0.0036460.193749 ± 0.008242
SWaT0.859716 ± 0.0097940.702476 ± 0.070891
Macro-average0.772067 ± 0.0094060.382516 ± 0.010517

We use the sample standard deviation, with denominator n−1=4. For the macro-average, we first average the five datasets with equal weights for each seed, then calculate the mean and standard deviation across the five seeds.


5. Memory size and computational cost

Regarding memory size, first-stage state-index search is affected by the size of the memory bank, whereas the candidate scope of second-stage local retrieval is bounded by a fixed number of windows. We fixed one checkpoint per dataset, reduced the state bank and the patch banks at both scales to 50% and 25%, and performed three timed runs after warmup on the same RTX 4090.

DatasetMemory ratioState windowsPatches per scaleAUROC / AP (%)Retrieval and distance scoring (ms/point, mean ± sample SD)Full scoring (ms/point, mean ± sample SD)
SMD100%708,278200,00082.50 / 19.420.16584 ± 0.000200.23375 ± 0.00072
SMD50%354,139100,00082.55 / 19.150.16496 ± 0.001880.23390 ± 0.00440
SMD25%177,07050,00082.76 / 19.260.16035 ± 0.001940.22923 ± 0.00279
SWaT100%494,873200,00087.19 / 63.890.20577 ± 0.001360.27269 ± 0.00160
SWaT50%247,436100,00087.14 / 65.210.20198 ± 0.002720.26903 ± 0.00356
SWaT25%123,71850,00086.72 / 69.660.19759 ± 0.005380.26389 ± 0.00660

Retrieval timing includes state search, candidate gathering, context matching, and distance scoring; full scoring additionally includes encoding, reconstruction, point-level aggregation, and fusion. Both use synchronized timing and exclude offline memory construction and calibration fitting.

Reducing memory to one quarter decreases retrieval time by 3.31% on SMD and 3.97% on SWaT, and full-scoring time by 1.93% and 3.23%, respectively. Stage-wise measurements show that state-search time decreases from 0.00228 to 0.00146 ms/point on SMD and from 0.00195 to 0.00134 ms/point on SWaT, while the main cost remains in subsequent local retrieval and scoring. Thus, within the measured range, larger memory does not lead to a proportional increase in retrieval time. Detection performance also does not vary monotonically with memory size: both metrics change little on SMD, while reducing memory on SWaT slightly decreases AUROC but increases AP, indicating that memory size should be chosen with both metrics in view.

The benefit of the additional computation is improved detection quality: in the original manuscript, SCAR improves macro-average AUROC by 10.88–22.84 percentage points and AP by 4.51–18.57 percentage points over these three baselines, at 9.1–12.7 times their inference cost. The performance gains and inference-time ratios below come from the original manuscript's performance table and Table 18, respectively.

BaselineSCAR inference-time ratioMacro-average AUROC gain (percentage points)Macro-average AP gain (percentage points)
PaAno12.7×+22.24+17.90
PUAD9.8×+10.88+4.51
PGRF-Net9.1×+22.84+18.57

SCAR primarily targets settings with complex, varying operating regimes, complex local temporal dynamics, and higher requirements for detection performance. In these settings, normal behavior must be judged in conjunction with operating state and local context, and the additional conditional retrieval computation constructs more suitable normal references; the detection gains above demonstrate the value of this computational investment. For settings with relatively stable operating regimes, simpler local dynamics, or greater emphasis on low latency, high throughput, and limited computational budgets, lighter detection methods should be considered.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.