acceptodds
Under review as a conference paper at ICLR 2027

Pretrain Once, Deploy Many Times: Frequency-Aware Memory for Training-Free Anomaly Detection on Unseen Machines

Abstract

Fault diagnosis foundation models promise reusable vibration representations, yet deployment still commonly requires target-specific training for each new machine. We introduce a pretrain-once, deploy-many-times framework: a vibration encoder is pretrained once on a heterogeneous cross-source corpus, frozen globally, and reused on unseen targets without target-domain gradient updates, while each target defines normality locally from a small set of normal references. FAMB (Frequency-Aware Memory Bank) instantiates this local detector through frequency-domain structure: it anchors matching to local Mel-frequency neighbourhoods and uses second-order latent frequency-shape residuals for bounded score enhancement. Same-sensor matching and unrestricted time positions preserve meaningful frequency comparisons. On four datasets excluded from pretraining, FAMB with one normal reference sample reaches 96.25%/96.92% macro-average AUROC/AUPR. On 16 liquid rocket engines, 3-fold device-level cross-validation keeps reference and test devices disjoint. With 100 samples from one normal engine, FAMB reaches 96.49%/96.88% AUROC/AUPR (across-fold standard deviation 1.98%/2.14%), exceeding the strongest baseline by 3.80 and 4.28 percentage points. These results support pretrain-once, deploy-many-times fault diagnosis through globally reusable vibration representations and locally defined normality.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
anomaly detection, self-supervised pretraining, representation learning, vibration signals, fault diagnosis, training-free adaptation, few-shot

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_FzFHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes FAMB, a frequency-aware memory bank for training-free anomaly detection on unseen machines. A vibration encoder is pretrained on 38 data sources using DINO-style self-distillation and frozen for deployment. FAMB builds a memory bank from normal reference samples and performs nearest-neighbor retrieval with sensor and frequency constraints while allowing flexible temporal matching. Second-order frequency-shape residuals further enhance anomaly scores. Experiments on four unseen vibration datasets and 16 liquid rocket engines evaluate cross-source transfer, few-shot detection, and cross-device generalization. The results show improvements over the implemented baselines, particularly when normal references are limited.

Critical strengths
  1. The paper addresses an interesting and practical problem: deploying anomaly detection models on unseen machines without target-specific training, particularly when only a few normal samples are available.

  2. The experiments are extensive, covering four unseen vibration datasets and 16 liquid rocket engines. The evaluation includes few-shot settings, cross-device transfer, and ablation studies, providing useful evidence of the method's effectiveness.

Critical weaknesses
  1. The baseline comparison is not sufficiently convincing. Most baselines are adapted from image anomaly detection methods, such as PaDiM [1], but the paper does not clearly explain how their image-specific components are adapted to vibration spectrograms. Meanwhile, relevant open-source signal anomaly detection methods, such as GenRep [2], are not considered, although differences in signal modality would need to be addressed.

  2. The methodological novelty appears limited. BEAM [3] has already explored frequency-local memory matching and training-free anomaly detection. Although FAMB introduces sensor-aware retrieval and second-order frequency-shape enhancement, the paper does not sufficiently clarify its contributions beyond these existing ideas.

[1] Defard et al., PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization. ICPR Workshops, 2021.

[2] Saengthong et al., Deep Generic Representations for Domain-Generalized Anomalous Sound Detection. ICASSP, 2025.

[3] Saengthong et al., Sub-Band Spectral Matching with Localized Score Aggregation for Robust Anomalous Sound Detection. arXiv, 2026.

Questions for authors
  1. Most detection baselines are originally designed for image anomaly detection. For example, PaDiM [1] relies on position-dependent Gaussian modeling of image features, which may not directly suit vibration spectrograms. I am concerned about how these image-specific components are adapted to signal anomaly detection, as the paper provides limited implementation details. Could the authors explain how each baseline is adapted, particularly its feature processing and anomaly scoring? Without appropriate adaptations, the comparison may not be fair, especially since FAMB is specifically designed for signal characteristics.

[1] Defard et al., PaDiM: a Patch Distribution Modeling Framework for Anomaly Detection and Localization. ICPR Workshops, 2021.

  1. Why are recent signal anomaly detection methods not included in the comparison? This field already has relevant open-source methods, such as GenRep [2] and PaAno [3]. Could the authors include these methods as additional baselines or explain why they cannot be fairly adapted to the evaluation datasets?

[2] Saengthong and Shinozaki, Deep Generic Representations for Domain-Generalized Anomalous Sound Detection. ICASSP, 2025.

[3] Park and Kang, PaAno: Patch-Based Representation Learning for Time-Series Anomaly Detection. ICLR, 2026.

  1. The main differences between FAMB and existing memory-based methods such as RAD appear to be frequency-constrained retrieval and second-order frequency-shape enhancement. Are these the main methodological contributions? Could the authors clarify what is fundamentally new beyond existing retrieval-based approaches? This is mainly a clarification request and does not affect my rating.
3 comments
AuthorsOpenReview ↗

Response to Reviewer FzFH

Thank you for recognizing the practical value of deployment with limited normal references. We clarify the baseline adaptations and FAMB's contribution below.

1. Adapting detection heads to vibration.

All heads use the same frozen vibration backbone, per-channel log-Mel inputs, normal references, test sets and repetitions. The “-style” labels identify unified implementations. Trainable heads optimize on target normal references while the encoder stays frozen (Appendix B.4).

HeadAdaptation and scoring
Global kNNNearest-neighbour comparison of global window descriptors.
PatchCore-style / RAD-stylePatch-memory retrieval / conditioned retrieval on shared frozen features.
PaDiM-styleA position is (sensor, frequency row, time column). A separate full Gaussian covariance is fitted at each position and regularized by 10−3I10^{-3}I. Mahalanobis distances are averaged over positions; the time axis is retained.
RD4AD-style / Dinomaly-styleTrain on normal features and score reconstruction discrepancy.
Deep SVDD-styleLearn a one-class boundary from normal-reference features.
BEAM-styleAverage and normalize features along time for each layer and sensor; retrieve independently in eight frequency sub-bands with reference-local density normalization; average sub-band, sensor and layer scores.

PaDiM therefore retains position-dependent modelling on spectrogram coordinates. FAMB instead preserves sensor and local-frequency correspondence while allowing time positions to vary with window boundaries. Table 3 directly tests these assumptions through controlled ablations.

2. Signal methods and the level of comparison.

GenRep uses generic pretrained representations and normal-reference retrieval for anomalous sound detection; PaAno learns patch representations with triplet and pretext objectives. Section 2 discusses both.

Our evaluation separates representation transfer from scoring with a shared frozen representation. Table E1 compares audio, time-series and vibration encoders under their documented interfaces. BEAM-style supplies the closest signal-domain head comparison: frequency-local memory matching using the same encoder and normal references as FAMB (Appendix B.4).

FAMB leads at all three reference budgets in Table 2:

Reference windowsBEAM-style AUROC / AUPR (%)FAMB AUROC / AUPR (%)
193.25 / 93.6896.25 / 96.92
295.91 / 96.4997.50 / 97.98
496.81 / 97.2998.40 / 98.65

At one window, the gains are 3.00 AUROC and 3.24 AUPR points, without added noise. Table C3 reports per-dataset means and SDs over 20 reference draws.

3. Contribution beyond BEAM and RAD.

The methodological contribution is the vibration-specific matching geometry and bounded second-order frequency-shape score (Section 3.2):

  • Frequency-local, time-free retrieval: FAMB retains individual time-position patches and matches within neighboring frequency rows at any time column. BEAM-style averages time before sub-band retrieval.
  • Sensor-conditioned normality: matches must share sensor identity. The engine study uses 11 consistent physical sensor positions, preventing features from a different sensing location from serving as normal matches.
  • Frequency-shape enhancement: second-order residuals measure departures in how normalized latent features vary across frequency, complementing nearest-neighbour coverage. The score is s=sbase(1+λa2)s=s_{\mathrm{base}}(1+\lambda a_2), with a2∈[0,1]a_2\in[0,1]; at λ=0.5\lambda=0.5, the multiplier is bounded to [1,1.5][1,1.5].

Table 3(b) isolates these effects with one reference engine. Freeing time matching raises AUROC from 92.58% to 94.21%; shape enhancement raises it to 96.49% and lowers FPR95 from 9.56% to 6.47%. Removing sensor conditioning or frequency anchoring reduces AUROC to 95.21% or 95.66%. These comparisons fix the backbone, reference choices and test windows, directly linking the gains to the matching and scoring design.

Thank you for the detailed response, and I apologize for my delayed reply due to other commitments. I no longer have major concerns about the fairness of the comparisons. My main concern remains the limited methodological novelty (W2). After reading the other two reviewers' comments, I noticed that they share similar concerns about the incremental nature of the contributions, although their views on whether these contributions and performance improvements are sufficient for acceptance differ. From my perspective, I would prefer to maintain my original score.

AuthorsOpenReview ↗

Response to Reviewer FzFH

Thank you for clarifying that W2 is your remaining concern. We agree that BEAM established frequency-local matching. When a new engine has no healthy baseline of its own, the matching rule determines which observations from another engine can serve as healthy references. FAMB preserves sensor location and local frequency correspondence while allowing the time offset within a recording window to vary.

A brief vibration at a pump sensor can fall in a different time column simply because the recording window starts later. Averaging over time before retrieval, as in our BEAM-style comparison, can dilute a brief departure. FAMB retains individual patches for matching across time. It uses second differences across neighboring frequency bands to describe the local shape of the learned feature profile. A patch mismatch receives more weight when that shape differs from those observed in healthy recordings.

The one-factor comparisons in Table 3(b) use the same frozen encoder, reference-engine choices and test windows at the one-engine budget. With the shape term off, allowing matches at any time offset raises AUROC by 1.63 points. Removing the frequency constraint from full FAMB lowers AUROC by 0.83 points. With retrieval fixed, adding the shape term reduces the false-positive rate at 95% sensitivity from 9.56% to 6.47%.

We would welcome your reassessment of W2.

Reviewer_t4uaOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper aims to reduce retraining when anomaly detection is deployed on a new machine. A vibration encoder is pretrained once and frozen.

FAMB defines local normality from a small reference memory using sensor/frequency-constrained retrieval and curvature-based scoring.

Tests include four unseen datasets and device-disjoint engines.

Critical strengths

The contribution is practically motivated: obtaining fault examples and retraining for each machine can be difficult, so adapting from a few normal references is useful. The retrieval rules also fit vibration data, where sensor identity and frequency location have meaning but exact time alignment may not. Device-disjoint evaluation gives the deployment claim substance: with one reference engine, FAMB exceeds PatchCore-style by 3.80 AUROC points. Component ablations support the design choices.

Critical weaknesses
  1. The novelty is modest. [1,2] already use frozen features and normal-reference memories. The contribution should therefore be judged as vibration-specific matching and scoring. The current ablations and device-disjoint results provide some support for that narrower contribution; a broader novelty claim is not necessary for acceptance.

  2. The pretraining comparison mixes method and data scale: 38 sources versus 400 target-normal windows. This limits attribution of the encoder gains, but does not invalidate the detection-head comparisons using a shared encoder. Likewise, four anomalous engines limit generalization claims. A larger engine study would be useful future work, rather than a requirement for rebuttal.

[1] Towards Total Recall in Industrial Anomaly Detection [2] Is Task-Specific Training Necessary for Anomaly Detection?

Questions for authors
  1. Can the authors clarify which gains they attribute to cross-source data and which to FAMB itself, using the shared-encoder comparisons? This would help establish the contribution without requiring a new pretraining study.

  2. How were the reduced reference-engine subsets selected, and do the existing per-device results indicate that the advantage is consistent rather than driven by a few devices?

2 comments
AuthorsOpenReview ↗

Response to Reviewer t4ua

Thank you for the thoughtful review and for recognizing the deployment motivation. We clarify the attribution of gains and the reference-engine protocol below.

1. Which gains come from cross-source pretraining, and which come from FAMB?

The two comparison groups answer separate questions:

  • Representation reuse: with the DINO objective, backbone, FAMB head and a 512,000-window pretraining budget fixed, Paderborn, four-source and 38-source pretraining yield 98.10%, 98.19% and 99.65% macro AUROC, respectively (Table E2). These results characterize the benefit of the tested cross-source pretraining configuration.
  • Local normality modelling: with the encoder weights fixed, changing only the detection head yields the following results (Tables 2 and D1):
Deployment settingStrongest compared baselineFAMBAUROC gain
One normal reference window, public datasetsRAD-style: 93.97%96.25%+2.28 points
One reference engine, device-disjoint test enginesPatchCore-style: 92.69%96.49%+3.80 points

These head-level gains do not involve different pretraining corpora or encoder weights. Table 3(b) further isolates the FAMB mechanisms at the one-engine budget: free time matching adds 1.63 AUROC points, shape enhancement 2.28, sensor conditioning 1.28 and frequency anchoring 0.83 in the corresponding one-factor comparisons.

Our methodological contribution is therefore the vibration-specific matching geometry and bounded second-order latent frequency-shape enhancement, within a reusable-encoder deployment framework. Section 2 explicitly identifies the inherited frozen-feature memory paradigm and distinguishes FAMB's matching and scoring from it.

2. How are the reduced reference-engine subsets selected?

Appendix E.2 specifies the selection within each device-disjoint fold. Its eight normal reference engines form the candidate pool; test engines and test windows remain fixed. Each selected engine contributes all 100 normal windows.

  • The eight-engine budget uses the full candidate pool.
  • For the two- and four-engine budgets, a preset random seed generates a permutation of the eight candidates. The first two or four engines are selected, making the subsets nested under the same seed.
  • For the one-engine budget, all eight candidate reference engines are evaluated in turn in every fold.

Every method uses identical reference engines, windows and test sets. Repeated runs are first averaged for each choice; choice means are then averaged equally within each fold. The reported mean and sample SD are calculated over the three fold means. Paired component gains use differences between matched configurations before the same within-fold aggregation.

3. Robustness to reference and test device variation.

At the one-engine budget, all eight candidate reference engines are evaluated exhaustively within each fold. FAMB's fold-averaged worst, median and best reference-choice AUROCs are 95.13%, 96.52% and 97.83% (Table E4). The matched component comparisons also yield positive mean paired gains after within-fold averaging.

Table E3 averages windows within each test engine before computing AUROC, giving one score per engine. FAMB reaches 95.83% machine-level AUROC versus 91.67% for PatchCore-style. The gain therefore persists at the engine level, while exhaustive reference selection avoids dependence on a single favorable reference engine.

Thanks to the authors who answered my main questions. I'd like to keep my score as it is.

Reviewer_Dh69OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a “pretrain once, deploy many times” framework for vibration-based anomaly detection for fault diagnosis on machines that were never seen during training. A small ViT is pretrained with DINO-style self-supervision on log-Mel spectrograms pooled from 38 public vibration datasets, and then kept frozen. To deploy on a new machine, the model simply stores features from a few normal samples in a memory bank, with no training on the target. The proposed detection head, FAMB, is a PatchCore-style nearest-neighbour retrieval adapted to vibration data: test patches are only matched against normal patches from the same sensor and nearby frequency bands, while time position is left unrestricted. It also adds a term that compares how features vary across frequency with normal behaviour, which boosts the anomaly score. The method is evaluated on four public datasets excluded from pretraining, in few-shot settings with as little as one normal sample, and on 16 liquid rocket engines where the normal references and test engines are physically different machines.

Critical strengths
  • The deployment setting is practical: no training per machine, only a few normal samples.
  • Evaluation boundaries are careful: target datasets are excluded from pretraining, records are split before windowing, and engines are device-disjoint.
  • Many detection heads are compared on the same backbone, so the head comparison is fair.
  • The few-shot and cross-device results are the most convincing (n = 1: 96.25 vs 93.97 AUROC for the best baseline and for 1 reference engine: 96.49 vs 92.69).
Critical weaknesses
  • The public benchmarks seem to be saturated.
    • Full-reference results are 99–100% AUROC; SEU-G is 100.00 for several heads, and even random weights reach 96.91 on it (Table C2).
    • Differences of 0.3–0.7 points can’t separate methods.
    • The authors add artificial noise (SNR 5 dB) to make it harder, which is an arbitrary choice.
  • The pretraining comparison is unbalanced. The single-source baselines are self-supervised ViTs trained on only 400 samples of the target, which is far too little data for these methods. So I think “Cross-source beats single-source” is mainly “more data beats very little data.”
  • Some relevant baselines are missing:
    • vibration foundation models they cite (UniFault, RmGPT, BearingFM, Mannone et al.);
    • time-series foundation models (e.g., Chronos, MOMENT);
    • general audio encoders (AST, BEATs, CLAP). The input is a log-Mel spectrogram, so I think AudioSet-pretrained models are the natural comparison.
  • The engine evaluation is statistically small.
    • There are only 16 engines, and only 4 anomalous, the same 4 in every fold.
    • AUROC is computed over windows, which are highly correlated within an engine, so window-level metrics overstate how reliable the results are. Engine-level detection results would be more meaningful.
  • The ablations are weak where the method claims its advantage.
    • The structural ablation (Table 3) is only on the engine data at the full 8-engine budget, where results are near saturation and gains are within the fold standard deviation (the authors admit this).
    • There’s no ablation in the few-shot or 1-engine setting, where FAMB’s advantage is largest.
    • There’s no sensitivity analysis for λ\lambda, τ\tau, ρf\rho f or top$_k$, and the paper doesn’t say how these values were chosen. Were they tuned on the engine data?
  • Possible overlap with pretraining. Several pretraining sources are bearing and gearbox datasets that may share test rigs or setups with the evaluation datasets. Exclusion is by dataset and record identity, so the authors should confirm there is no rig-level overlap.
  • The encoder is pretrained with one seed (42), so there is no variance over pretraining.
  • Memory size and inference cost are not reported.
Questions for authors

I would like to hear the authors' comments on the weaknesses I raised above, particularly around:

  1. cross-source encoder comparison with existing pretrained encoders, including vibration foundation models, time-series foundation models, and audio encoders such as AST or BEATs
  2. How were the FAMB hyperparameters (λ\lambda, τ\tau, ρf\rho f and top$_k$) selected, and how sensitive are the results to them?
  3. Structural ablation in the few-shot and 1-reference-engine settings, where FAMB’s advantage is claimed
  4. How about the engine-level evaluations in addition to the window-level evaluations (given that windows from the same engine are correlated)?
  5. Do any pretraining sources share test rigs or setups with the evaluation datasets?
4 comments
AuthorsOpenReview ↗

Response to Reviewer Dh69 (1/2)

Thank you for the careful review and for recognizing the practical setting and shared-backbone comparisons. We address the experimental concerns below; protocol and implementation details follow in Part 2.

1. Benchmark saturation and noise.

Our strongest deployment evidence uses no added noise. With one reference window, FAMB reaches 96.25% AUROC versus 93.97% for the strongest baseline; with one reference engine and physically disjoint test engines, it reaches 96.49% versus 92.69% (Tables 2 and D1).

Noise robustness is relevant to industrial monitoring because measurement disturbances can mask fault signatures. Multi-SNR Gaussian-noise testing is established in bearing diagnosis, e.g., Nguyen et al., 2021, Section 4.5. Our 5 dB setting is an intermediate point in the 0, 5, 10 and 20 dB sweep (Table D3). FAMB leads at 0, 5 and 10 dB, with 98.34%, 99.65% and 99.94% AUROC; at 20 dB, competitive heads approach the ceiling. The 2.28- and 3.80-point deployment gains therefore stand independently of the noise setting.

2. Pretraining data volume.

Table E2 compares three configurations using exactly 512,000 distinct windows, sampled without replacement, separately from the 400-target-window control. The DINO objective, backbone, FAMB head and 20-epoch schedule are fixed:

Pretraining sourcesWindowsAUROC / AUPR (%)
Paderborn512,00098.10 / 98.17
Four sources512,00098.19 / 98.25
38 sources512,00099.65 / 99.66

This supports broader source coverage at the tested budget. For the detection-head comparisons, all methods share the same frozen encoder, isolating FAMB's contribution to local normality modelling (Sections 4.2–4.4).

3. External pretrained representations.

Table E1 includes AST, BEATs, CLAP, Chronos, MOMENT, UniFault and RmGPT. Audio and time-series models use official checkpoints; the vibration entries use their published architectures and objectives trained on the same 38-source corpus. BEATs reaches 97.76% macro AUROC and the UniFault implementation 97.89%, compared with 99.65% for our encoder.

Appendix E.1 specifies checkpoints, read-out positions, native dimensions and feature grids. CLAP uses global matching with the shape branch disabled; RmGPT uses the documented fixed read-out. These are results under the stated adaptation interfaces, with the vibration entries representing same-corpus implementations. The shared-encoder head comparison separately controls representation differences.

4. Ablation with one reference engine.

Table 3(b) uses one reference engine / 100 normal windows, with identical reference choices, test windows and seeds across configurations:

ConfigurationAUROC (%)FPR95 (%)
Frequency-local, time-free retrieval without shape94.21 ± 2.089.56 ± 0.72
Full method without sensor constraint95.21 ± 2.038.52 ± 0.58
Full method without frequency anchoring95.66 ± 2.067.56 ± 0.69
Full FAMB96.49 ± 1.986.47 ± 0.45

Shape enhancement adds 2.28 AUROC points and reduces FPR95 by 3.09 points. Free time matching adds 1.63 AUROC points over isotropic conditioning. Table E4 reports paired gains of 2.28 ± 0.36, 1.28 ± 0.32 and 0.83 ± 0.29 points for shape enhancement, sensor conditioning and frequency anchoring. Each gain is averaged over matched choices within folds before reporting the mean and sample SD across folds.

For the separate single-window setting (Table 2), shape statistics use the reference's eight time columns with leave-one-column distances and shrinkage (Section 3.2). These estimate short-time variation without increasing the number of reference samples.

AuthorsOpenReview ↗

Response to Reviewer Dh69 (2/2)

  1. Engine-level evaluation.**

Table E3 assigns one score to each engine, averaging its window scores before computing AUROC over four normal and four anomalous test engines per fold. With one reference engine, FAMB reaches 95.83 ± 3.61%, compared with 91.67 ± 3.61% for PatchCore-style. FAMB also leads at the other three budgets, so its advantage remains after accounting for within-engine window correlation.

The four anomalous engines are test-only, and normality is constructed from physically separate normal engines. This directly tests an industrial deployment need: assessing another hardware unit using available normal records, without target fault examples for detector construction or target-machine gradient updates. All eight eligible reference engines are evaluated in turn at the one-engine budget.

Appendix E.2 first averages repeated runs for each reference choice, then choices within each fold, and reports the mean and sample SD of the three fold means. With the same four anomalous engines held fixed, this SD describes fold variation.

6. Hyperparameter settings and sensitivity.

The fixed defaults have a direct interpretation (Section 3.2; Appendix B.3):

  • ρf=1\rho_f=1: match the query's frequency patch row and its immediate neighbors.
  • λ=0.5\lambda=0.5: bound the score multiplier to [1,1.5][1,1.5].
  • τ=1.5\tau=1.5, γ=0\gamma=0: control the enhancement response and set its onset at the estimated normal residual mean.
  • topk=8\mathrm{top}_k=8: retain eight reference windows, or all when fewer are available.

These defaults were fixed before the reported comparisons and kept constant across datasets, devices and budgets. Test windows and anomalous samples do not enter parameter selection. Table E5 is a post-hoc, one-parameter-at-a-time check.

At the one-engine budget, the full tested ranges of ρf\rho_f, τ\tau and γ\gamma give 96.41–96.49%, 96.18–96.49% and 96.38–96.49% AUROC. Their largest range is 0.31 points, compared with the 3.80-point default-setting margin over PatchCore-style. Tested λ\lambda values from 0.25 to 1.0 give 96.18–96.49%; λ=0\lambda=0 gives 94.21%, matching the shape ablation. Across all tested topk\mathrm{top}_k settings, AUROC is 95.54–96.49%, with 8, 16 and 32 yielding 96.49%, 96.45% and 96.28%. These results support a useful operating region around the defaults.

7. Pretraining and evaluation separation.

All four public evaluation datasets and their raw records are excluded from pretraining. Duplicate or mirrored records are removed by record identity, and reference/test records are split before windowing (Section 2; Appendix A). This establishes unseen-source transfer at dataset and record level; physical-rig identity is a separate granularity.

Physical-device transfer is tested by the engine study: all 16 engines are excluded from pretraining, and reference and test engines are disjoint within each fold. The public and engine experiments thus provide complementary evidence at their respective levels of separation.

8. Pretraining seeds.

Seeds 42, 123 and 2024 yield 99.65%, 99.58% and 99.61% macro AUROC. At one reference window, FAMB's macro AUROC varies by 0.28 points across seeds, compared with its 2.28-point margin over the strongest baseline (Section 4.3; Appendix E.1). Within each head comparison, all methods share the same frozen weights.

9. Memory and scoring cost.

Table D2 measures detection heads on an A100, using float16 memories and batch size 64. At one reference engine, FAMB stores 144.2 MB of patch features across 11 sensors, with 0.13 s for memory construction and 4.31 s for scoring 800 test windows. At eight engines, FAMB, PatchCore-style and RAD-style each store 1,153.4 MB, with scoring times of 4.47 s, 3.62 s and 3.05 s. These timings use cached features and exclude encoder extraction. Deployment cost depends on the reference budget, not the size of the pretraining corpus.

I thank the authors for their effort in providing more results and responding to my and other reviewers' concerns and questions. Their responses answer some of my concerns/questions, while others are not answered adequately. After reading fellow reviewers' comments and the authors' responses, I have decided to leave the score unchanged, particularly because of the limited novelty and insignificant performance gain, as the public benchmark seems to be saturated.

AuthorsOpenReview ↗

Response to Reviewer Dh69

Thank you for following up. We share your caution about the near-ceiling full-reference scores on SEU-G. In the single-window evaluation on that dataset, mean AUROC over 20 reference draws is 98.22% for FAMB and 95.60% for RAD-style, the next-best head (Table C3). FAMB can use a healthy pattern from that one window even when it occurs at a different time offset, while preserving its sensor and frequency context.

In the engine study, the largest gain occurs in our target setting, where only one healthy reference engine is available. With eight reference engines per fold, FAMB leads PatchCore-style by 0.33 points in mean window-level AUROC. When each of the eight serves in turn as the sole reference, both methods use the same 100 healthy windows, frozen encoder and held-out test engines. FAMB reaches 96.49% window-level AUROC and PatchCore-style reaches 92.69%, a 3.80-point gain (Table D1).

Your original review also asked how the methods compare when each test engine counts once. Averaging the window scores within each test engine gives 95.83% machine-level AUROC for FAMB and 91.67% for PatchCore-style at the one-engine budget (Table E3). In this setting, a team uses healthy recordings from one unit to assess other engines.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response: Revision Summary and Shared Concerns

We thank all three reviewers for their constructive feedback and for recognizing the practical value of deploying anomaly detection with limited normal references. The revised manuscript strengthens the evidence for this setting and clarifies the contribution and evaluation protocol.

1. What changed in the revised manuscript

  • Contribution and related work: Section 2 clarifies FAMB's relation to PatchCore, RAD, BEAM, GenRep and PaAno. The contribution centers on vibration-specific matching and bounded second-order frequency-shape enhancement within a reusable-encoder framework.
  • Baseline implementation: Appendix B.4 details the spectrogram adaptations, including PaDiM's sensor–frequency–time positions and BEAM-style sub-band retrieval. Appendix E.1 documents external encoders' checkpoints, feature read-outs and adaptation interfaces.
  • Engine protocol and statistics: Section 4.1 and Appendix E.2 specify the seeded two-/four-engine subsets, exhaustive single-engine choices, shared references across methods, and within-fold averaging before reporting the mean and sample SD over three folds.

2. Added experiments and tables

Added or extended evidenceLocation
BEAM-style comparisons at full and 1/2/4-window reference budgetsTables 2, C2–C3
Seven external audio, time-series and vibration encodersTable E1
Pretraining comparisons at a matched budget of 512,000 distinct windowsTable E2
Structural ablation with one reference engine, alongside the eight-engine settingTable 3(b)
Engine-level AUROC using one score per test engineTable E3
Exhaustive single-reference-engine spread and paired component gainsTable E4
Parameter sensitivity at one- and eight-engine budgetsTable E5
Detection-head memory and timing from cached features; noise sweep at 0/5/10/20 dBTables D2–D3
Three pretraining seeds, including the single-window settingSection 4.3; Appendix E.1

3. Responses to concerns shared across reviews

Attribution of gains and comparison fairness (Dh69, t4ua, FzFH). We separate representation learning from target normality modelling. Table E2 fixes the DINO objective, backbone, detector, 20-epoch schedule and 512,000-window budget while varying source coverage: one, four and 38 sources yield 98.10%, 98.19% and 99.65% macro AUROC. Separately, detection heads share identical frozen encoder weights and reference sets. Thus, FAMB's head-level gains are evaluated without differences in pretraining data access. BEAM-style provides a signal-domain head comparison; Table E1 addresses external representation transfer under documented interfaces. Audio/time-series entries use official checkpoints, while vibration entries use the published architectures and objectives trained on our corpus.

Contribution beyond existing memories (t4ua, FzFH). FAMB preserves sensor identity and local frequency correspondence while allowing free temporal matching, then supplements retrieval with bounded latent frequency-shape residuals. Table 3(b) isolates these choices under limited normal coverage: at one reference engine, free time matching adds 1.63 AUROC points and shape enhancement adds 2.28 points, reducing FPR95 from 9.56% to 6.47%. Removing sensor conditioning or frequency anchoring also reduces performance. These results connect the proposed design to the deployment problem.

Device-level reliability and reference choice (Dh69, t4ua). The single-engine evaluation uses all eight eligible reference engines in each fold, with identical choices and test sets across methods. Table E3 gives each test engine one score before computing AUROC: FAMB reaches 95.83%, compared with 91.67% for PatchCore-style. The advantage therefore remains at the engine level. The reported SD describes variation across fold means with the same four anomalous engines held fixed.

The central deployment results use no added noise: FAMB exceeds the strongest compared head by 2.28 AUROC points with one normal window and 3.80 points with one reference engine. Together with the added controls and engine-level analysis, these results support transferring normality across machines without target-domain gradient updates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.