acceptodds
Under review as a conference paper at ICLR 2027

When Labels Are Scarce: An Oscillatory State Space Model for Vibration Diagnosis

Abstract

Machine fault diagnosis from vibration requires learning from scarce labelled fault recordings while meeting the computational constraints of edge devices for local inference. We introduce DualRes, a compact oscillatory state-space model that combines two complementary spectral views of vibration, capturing rapid changes and fine frequency structure. Time-aligned views are processed by selective oscillatory memory, which learns how long to retain temporal patterns. The encoder contains 39,528 parameters. We evaluate supervised learning across six bearing datasets and a gearbox benchmark, with an additional gearbox pilot. Recording-level splits and explicit accounting of labelled duration distinguish data efficiency from repeated exposure to correlated samples. On the main gearbox benchmark, DualRes achieves state-of-the-art performance among the nine evaluated methods at six of seven label budgets. With about six labelled seconds per class, it improves macro-F1 by 16.1 percentage points over the next strongest comparator. On the same benchmark, DualRes achieves a 1.44-fold recording-level speedup and a 24.8-fold reduction in checkpoint storage relative to a selective state-space baseline under matched hardware and runtime conditions. However, performance on bearing datasets varies across tasks. These findings support oscillatory memory as a compact approach to vibration diagnosis under limited labelled exposure.

Primary area:
learning on time series and dynamical systems
Keywords:
state space models, vibration-based fault diagnosis, data-efficient learning, oscillatory dynamics, multi-resolution spectral representations, efficient inference

Reviews

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

The paper introduces DualRes, a compact vibration classifier that combines two time-aligned STFT resolutions with selective oscillatory memory. The model is evaluated under limited labeled-data budgets using recording-level train/test splits. The paper also reports component ablations and analyses of computational efficiency and memory usage.

Critical strengths
  1. The recording-level splits and explicit accounting of labeled duration make the scarce-label evaluation more credible.
  2. The compact architecture is well motivated for vibration signals. The no-rotation ablation supports the usefulness of oscillatory dynamics.
Critical weaknesses
  1. The component ablations use only one seed, despite substantial seed variation on some tasks. This makes the size and consistency of the component gains uncertain. Besides, jitter has little apparent benefit. Removing it leaves most results unchanged. Its intended role and the reason for retaining it are not clearly justified.
  2. Table 2 covers seven label budgets, but several baselines remain weak even at the largest budget. It is unclear whether this mainly reflects label scarcity, the training settings, or difficulty handling the domain shift. This makes it harder to attribute the performance gap specifically to better data efficiency.
Questions for authors
  1. Could you repeat the key ablations over multiple seeds and report the mean and variability of the gains?
  2. What is damping jitter intended to improve, and why retain it given the limited gains in Table 4?
  3. How do you interpret the weak baseline results at the largest label budget in Table 2? Can the available results distinguish label scarcity from training settings or domain shift?
5 comments
AuthorsOpenReview ↗

Added five-seed ablations, jitter justification and training-setting and motion-transfer studies

Response to Reviewer 3

Thank you for the assessment. We have expanded the ablations and added studies of training settings and operating conditions. Locations refer to the revised paper.

Responses to the weaknesses

W1. Seed variation and the role of damping jitter

We repeated the full model and all four controls across five seeds on all seven datasets. The controls use short-only inputs, long-only inputs, zero rotation or zero jitter. Each seed averages both predefined folds. Table 5 reports the mean and sample SD, and Table 16 reports paired full-model-minus-control differences.

Rotation improves mean macro-F1 on six of seven datasets, with effects from −2.18 to +9.85 percentage points. The resolution controls also show dataset-dependent effects. Short-only inputs perform best on Paderborn, while long-only inputs lead on KAIST and MCC5.

We have also explained jitter's purpose and reported its gains and losses by dataset.

Changes Section 3.3 (p. 5), Section 5.4 and Table 5 (pp. 8–9), Appendix H and Table 16 (pp. 29–30).

W2. Weak baselines at the largest label budget

We added two diagnostic studies to examine training settings and transfer between motion modes. The training-setting study tests three learning rates and 300/600 updates for five models. Settings are selected using held-out source recordings before target evaluation.

The motion study trains each model on three recordings/class. It then tests on separate recordings from either the same motion mode or the other mode. The test recordings are matched for class, fault severity, nominal speed and load. Both diagnostics use seed 41 and average the two predefined folds.

Figure 5 now includes all ten neural methods. Several comparators reach near-zero training loss yet obtain lower test macro-F1. This shows that low training loss alone gives an incomplete account of the performance gap. The added diagnostics examine how the gap changes with training choices and test conditions.

Changes Section 5.1 (pp. 6–7), Section 5.5 and Figure 5 (p. 9), Appendices G.3–G.4 and Tables 14–15 (pp. 27–29).

Answers to the questions

Q1. Multi-seed ablations with mean gains and variability?

Yes. Tables 5 and 16 now cover five matched seeds for every component control on all seven datasets. We calculate each paired effect after averaging the two folds within a seed.

For rotation, the gain is 9.85 ± 11.91 points on UORED, 6.67 ± 2.42 on HUST and −2.18 ± 10.50 on MCC5. These values show substantial variation on UORED and MCC5. Pairing the runs measures variation in the component's effect across seeds.

The ZIP includes the ablation configurations, seed assignments and commands that calculate the paired effects.

Locations Section 5.4 and Table 5 (pp. 8–9), Appendix H.3 and Table 16 (p. 30), Appendix J and Table 20 (pp. 32–34). ZIP details appear in docs/RUNS.md.

Q2. What is jitter intended to improve, and why retain it?

We welcome feedback on these revisions.

Damping jitter is intended to reduce dependence on one precise memory-decay rate. During training, we sample a positive damping multiplier once per window and channel. The multiplier is shared across time and has an expected value of one. We use the same strength, σ = 0.10, across datasets and disable jitter at inference.

The five-seed paired gains are 6.24 ± 3.47 points on UORED and 4.34 ± 0.92 on HUST. CWRU and Ottawa have zero change, while MCC5 changes by −0.91 ± 1.83 points. Thus, the strongest evidence for retaining this training regularizer comes from UORED and HUST. We keep one fixed configuration across datasets and report the effect on each task.

The no-jitter control preserves the random draw used by jitter, so subsequent dropout draws stay aligned.

Locations Section 3.3 (p. 5), Section 5.4 (p. 9), Appendix H.2–H.3 and Table 16 (p. 30).

Q3. Can the results distinguish data scarcity, training settings and domain shift?

In the motion study, increasing labelled duration from 6.144 to 36.864 seconds/class raises within-motion macro-F1 from 5.31% to 31.46% for PatchTST. The recording count stays fixed, although the additional windows also increase computation per update.

In the training-setting study, source-validation selection raises MSCResNet from 46.55% to 57.49% at 36.864 seconds/class. DualRes changes from 65.96% to 67.29%. At this budget, we retrain each model on the full labelled pool using its selected settings. After this selection, DualRes retains a 9.80-point macro-F1 advantage over MSCResNet at the largest budget.

In the motion study at the same duration, MSCResNet scores 58.66% within the training motion and 35.48% across motions. DualRes scores 68.60% and 72.84%, respectively. Thus, the effect of changing motion depends on the model. These results link the observed gaps to labelled duration, training settings and motion conditions.

Locations Section 5.1 (pp. 6–7), Appendices G.3–G.4 and Tables 14–15 (pp. 27–29).

We would welcome your feedback on whether these revisions address your concerns.

Thank you for the additional experiments. They address my main concerns, and I’m willing to raise my rating to 3.

Reviewer_as7SOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies vibration diagnosis with very limited labeled data. It proposes DualRes-OSSM. The model combines two STFT resolutions with an oscillatory state-space layer. The experiments cover several bearing and gearbox datasets. The results show that the model is competitive under small label budgets while using relatively few parameters.

Critical strengths

1.The experimental protocol is careful. The data are split by recording or physical unit before windowing, which avoids a common form of leakage in vibration diagnosis. The label budget is also defined by signal duration, making the low-label setting more meaningful.

2.The method in this paper is compact, but it achieves strong or competitive results across several bearing and gearbox datasets and under different label budgets.

3.The model design fits the problem well. The oscillatory state-space layer provides a reasonable inductive bias for vibration signals, while the dual-resolution frontend captures both short- and long-term spectral patterns. The released code and fixed experimental splits also make the work easier to examine and reproduce.

Critical weaknesses

The main improvement is not fully isolated. DualRes-OSSM uses a dual-resolution spectral frontend, while most baselines use raw signals or their own frontends. A comparison that keeps the same frontend and only changes the temporal model is missing. It is therefore difficult to know how much of the gain comes from the oscillatory state-space layer. The results are not equally stable across all datasets. Some components help on one dataset but not on another. Only a few methods are repeated with multiple seeds. The released code also has not been used to rerun the complete experimental campaign. These issues reduce my confidence in the exact size of the reported improvement.

Questions for authors
  1. Can the authors compare DualRes-OSSM with another temporal model using exactly the same dual-resolution STFT frontend and training protocol?

  2. How were the baseline hyperparameters selected? Were they tuned separately for each dataset, or were the reported configurations fixed before seeing the test results?

  3. Can the authors provide multi-seed results for more of the baselines, especially on the datasets where the improvements are small?

5 comments
AuthorsOpenReview ↗

Added a shared-frontend comparison, broader five-seed results and a complete campaign rerun

Response to Reviewer 2

Thank you for the assessment. We have strengthened the comparisons and reproducibility. Locations refer to the revised paper.

Responses to the weaknesses

W1. Separating frontend and temporal-model gains

We added a comparison between oscillatory memory and a standard Mamba layer. Both models use the same dual-resolution STFT frontend architecture, pooling, classifier, labelled windows and training protocol. Each model learns its frontend parameters independently.

Across five seeds, DualRes has higher mean macro-F1 on all seven datasets. On MCC5, it scores 47.75 ± 7.61%, compared with 43.12 ± 4.64% for same-STFT Mamba. On Paderborn, the difference is smaller, at 49.73 ± 1.75% versus 48.17 ± 0.79% (Table 4).

Holding the frontend architecture fixed gives direct evidence about the contribution of the temporal model. The results also show how the size of this contribution varies across tasks.

Changes Section 5.3 and Table 4 (p. 8), Appendix D.3 (pp. 21–22).

W2. Dataset-dependent effects and limited seed coverage

The primary comparison now uses five seeds for all ten neural methods and MiniRocket. Each seed averages both folds before we calculate the mean and sample SD. The shared-frontend comparison and all four component ablations also use five seeds.

The revised discussion describes where each component helps. Rotation improves mean macro-F1 on six of seven datasets, with changes from −2.18 to +9.85 points. Short-only inputs perform best on Paderborn, while long-only inputs lead on KAIST and MCC5. Appendix H reports paired full-model-minus-control differences, calculated within each seed, and their sample SDs.

Thus, readers can assess both the average component effect and its variation.

Changes Section 4 (pp. 5–6), Section 5.1 and Table 1 (p. 6), Sections 5.3–5.4 and Tables 4–5 (pp. 8–9), Figure 9 and Tables 11–13 (pp. 22–25), Appendix H (pp. 29–30).

W3. Reproduction through the released code

We reran the complete experimental campaign through the packaged entry points. Appendix J.3 now records this completed rerun. The updated ZIP includes training and evaluation entry points for all twelve methods, the shared-frontend comparison, component ablations and the diagnostic studies. It also provides predefined splits, window allocations, seed settings and aggregation commands.

The package checks cover raw-data conversion, adapter outputs, gradients and checkpoint loading. The added neural adapters and controls agree with their research implementations to within 5 × 10⁻⁷ on the checked inputs. Bispectral features match exactly.

Changes Reproducibility Statement (p. 11), Appendix J and Table 20 (pp. 32–34). ZIP documentation includes docs/RUNS.md, docs/VERIFICATION.md and PACKAGE_CHECKS.json.

Answers to the questions

Q1. Same frontend and training protocol with another temporal model?

Yes. The new control uses a standard Mamba layer with the same 256/1024-sample STFT windows, hop of 128 and aligned 249-frame sequence as DualRes. Pooling, classifier, label budgets and training settings are shared. Both models use the final checkpoint. Table 4 reports five-seed results for all seven datasets.

Locations Section 5.3 and Table 4 (p. 8), Appendix D.3 (pp. 21–22).

Q2. How were baseline hyperparameters selected?

The broad comparison uses one documented architecture configuration per neural method across datasets, with a task-specific output head. Table 10 and Appendix D give these settings and adaptations. All neural models use Adam at 5 × 10⁻⁴, with decay of 0.99 per update. Bearing runs use 100 updates. The source-validation pilot selected the shared 300-update MCC5 duration.

We also added a separate source-validation study for five models. It tests three learning rates and 300/600 updates. For each model, source fold and budget, we select settings using validation macro-F1 before target evaluation. At the primary budget, this uses an additional 6.144 validation seconds/class. The study uses one seed and averages both folds.

The reported runs use these fixed configurations and final checkpoints. Section 4 records repeated development on these benchmarks, which limits independent confirmation.

Locations Section 4 (pp. 5–6), Appendix C.3 (p. 19), Appendix D and Table 10 (pp. 20–22), Appendices G.2–G.3 and Table 14 (pp. 26–28).

Q3. More multi-seed baselines where differences are small?

Yes. Table 12 covers all twelve methods at the primary budgets. The ten neural methods and MiniRocket have five-seed summaries. Spectral–SVM uses its deterministic procedure. For example, CWRU scores are 67.98 ± 2.99% for DualRes and 65.39 ± 7.71% for MambaSL. The small mean gap and variation are now visible together.

Locations Tables 1 and 12 (pp. 6, 23–25), Figure 9 (p. 22), Appendix E.4 (pp. 23–24).

We welcome feedback on these revisions.

We would welcome your feedback on whether these revisions address your concerns.

Thanks for your rebuttal, my main concerns are solved. I therefore keep my recommendations for accept.

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

Research paper proposes DualRes - efficient model for machine vibration signal fault diagnosis in case when very few labels are available. DualRes combines short and long STFT views of signal with selective oscillatory state space memory and has about 40K parameters. It is evaluated on six bearing datasets and gearbox benchmark.

Critical strengths
  1. Good applied problem statement. Fault diagnosis with very few labels and limited hardware resources is good real life practical problem.

  2. Hybrid machine learning model combined with mathematical modeling. Model design is based on vibration modeling from physics with two spectral resolutions, damped oscillatory memory that keeps model very small.

Critical weaknesses
  1. Limited novelty - main concern. Problem statement and datasets used are standard, many existing research papers study fault diagnosis on these benchmarks. Machine learning model combines existing parts: multi-resolution spectral inputs, damped rotating states.

  2. Contribution is not enough for ICLR. Research contribution is mainly applied to vibration data with no big insights for wide community of ICLR. This research work is good and practical, but usually published at conferences focused on fault diagnosis or signal processing.

  3. Baselines are weak. Most baselines are general time series models and several of them show insignificant improvement on gearbox benchmark. This could mean that they are not tuned for this task. Simple fault diagnosis baselines such as CNN on spectrograms or classical spectral features, are missing.

  4. Results on bearing datasets are mixed. DualRes does not show improvement on half of bearing datasets. Variance of results is high, only three seeds are used.

Questions for authors
  1. Did you compare DualRes with standard fault diagnosis baselines (for example, small CNN on spectrograms)?

  2. Did time series baselines used same spectral input and training parameters as proposed method?

7 comments
AuthorsOpenReview ↗

Clarified architectural novelty and added controlled comparisons, fault-diagnosis baselines and five-seed results

Response to Reviewer 1

Thank you for the assessment. We have strengthened the comparisons. Locations refer to the revised paper.

Responses to the weaknesses

W1. Limited novelty

Section 2 relates DualRes to LRU's input-independent complex transitions and Mamba's selective updates. DualRes combines input-dependent retention with separately learned rotation, so memory duration can change independently of phase advance. Each input write is scaled by one minus the retention coefficient. Appendix A.4 expands the recurrence to show how retention, rotation and input writes determine the contribution of each past input.

We also added a controlled temporal-model comparison. Both models use the same dual-resolution spectral frontend architecture, pooling, classifier and training protocol. Across five seeds, DualRes has higher mean macro-F1 on all seven datasets. On MCC5, it scores 47.75 ± 7.61%, compared with 43.12 ± 4.64% for same-STFT Mamba (Table 4).

This comparison supports a contribution from the memory formulation with the frontend architecture held fixed.

Changes Sections 1–3 (pp. 1–4), Section 5.3 and Table 4 (p. 8), Appendix A.4 (p. 16).

W2. Relevance to machine learning

We examine how temporal memory and recording diversity affect learning from few labelled recordings. Five-seed ablations show that rotation improves mean macro-F1 on six of seven datasets, with effects from −2.18 to +9.85 points. At a fixed duration, DualRes scores 47.75% with four windows from each of three recordings, compared with 25.67% for one recording and 44.20% for six. These controls connect model design with data efficiency.

A small number of recordings can produce many correlated windows. Equal labelled duration can therefore give different results when recording coverage changes. This supports reporting both duration and recording coverage in scarce-label benchmarks.

Changes Section 1 (pp. 1–2), Sections 5.2–5.4 and Figure 4 (pp. 7–9), Tables 4–5 (p. 8), Section 6 (p. 10).

W3. Missing baselines and training settings

The comparison now includes twelve methods. We added MSCResNet (Xin et al., 2022), TFN-STTF (Chen et al., 2024a), and Spectral–SVM (Saidi et al., 2015). Appendix D gives their adaptations. At 6.144 labelled seconds/class on MCC5, their macro-F1 scores are 35.69 ± 3.83%, 26.40 ± 4.55%, and 25.42%, respectively. DualRes scores 47.75 ± 7.61% (Table 12).

Appendix G.3 tests three learning rates and 300/600 updates for five models. Settings are selected using source-validation data. At 36.864 seconds/class, MSCResNet improves from 46.55% to 57.49%, while DualRes changes from 65.96% to 67.29%. This study uses one seed and averages both folds. At the primary budget, selection uses 6.144 additional validation seconds/class. Figure 5 adds the two neural baselines' loss curves, and Appendix G.4 tests changes in motion mode.

The selected settings reduce the performance gap, while DualRes retains the highest score among the five models. Several comparators also reach low training loss but score below DualRes on test data (Figure 5).

Changes Section 2 (p. 2), Section 5.1 (pp. 6–7), Figure 5 (p. 9), Appendix D (pp. 20–22), Tables 12–13 (pp. 23–25), Appendices G.3–G.4 and Tables 14–15 (pp. 27–29).

W4. Mixed bearing results and seed variation

The primary comparison uses five seeds for all ten neural methods and MiniRocket. Each seed averages both folds before we calculate the mean and sample SD. Shared-frontend comparisons and component ablations also use five seeds.

Against MambaSL, DualRes now has higher means on four bearing datasets, an equal mean on Ottawa, and a lower mean on HUST. UORED still has substantial variation (65.87 ± 12.21%), and MiniRocket leads on HUST. Tables 1 and 12 and Figure 9 report these results.

Appendix H also reports paired differences to show how component effects vary across seeds.

Changes Section 4 (pp. 5–6), Section 5.1 and Table 1 (p. 6), Figure 9 and Appendix E.4 (pp. 22–24), Tables 11–12 (pp. 23–25), Appendix H (pp. 29–30).

Answers to the questions

Q1. Comparison with fault-diagnosis baselines?

Yes. MSCResNet uses a spectrogram CNN, TFN-STTF learns time-frequency filters, and Spectral–SVM uses bispectral features. Table 10 gives their settings.

Locations Table 10 (p. 21), Table 12 (pp. 23–25).

Q2. Same spectral inputs and training parameters?

The broad comparison uses each method's documented input representation, with the same labelled recordings and windows. Neural models use Adam, an initial learning rate of 5 × 10⁻⁴, decay of 0.99 per update, and 300 MCC5 or 100 bearing updates. Normalization and augmentation follow each implementation.

The added same-STFT Mamba control matches DualRes's frontend architecture and training protocol. Each model learns its frontend parameters independently.

Locations Appendix C.3 (p. 19), D.1–D.3 (pp. 20–22), Section 5.3 and Table 4 (p. 8).

We welcome feedback on these revisions.

We would welcome your feedback on whether these revisions address your concerns.

Reviewer_UnKNOpenReview ↗

Comment for "Clarified architectural novelty and added controlled comparisons, fault-diagnosis baselines and five-seed results"

Thank you for providing the answers. All but novelty explainations and additional results are accepted. However, at this moment from my side the novelty level is not enough to be published in ICLR conference, but taking into account other reviewers I would like to continue the discussion.

Please use only 1-2 sentences to explain again and clearly what is the novelty of your method and how it could potentially be applied to a wider class of problems, not only to Vibration Diagnosis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.