pith. sign in

arxiv: 2604.19658 · v1 · submitted 2026-04-21 · 💻 cs.LG

Disentangling Damage from Operational Variability: A Label-Free Self-Supervised Representation Learning Framework for Output-Only Structural Damage Identification

Xudong Jian , Charikleia Stoura , Simon Scandella , Eleni Chatzi This is my paper

Pith reviewed 2026-05-10 03:44 UTC · model grok-4.3

classification 💻 cs.LG
keywords disentangled representation learningself-supervised learningstructural health monitoringvibration-based damage identificationlabel-freeautoencoderoperational variabilityVICReg
0
0 comments X

The pith

A self-supervised autoencoder framework learns to separate structural damage signals from operational and environmental variability using only unlabeled vibration data.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper presents a label-free self-supervised framework that trains an autoencoder directly on raw acceleration signals to produce a representation sensitive to damage but invariant to changes in excitation or environment. One latent code is regularized with VICReg on baseline recordings assumed to have fixed damage, while a power spectral density constraint forces the representation to retain frequency-domain agreement with the input. This combination allows end-to-end training without damage labels or knowledge of conditions, and the resulting representation supports both detection and quantification tasks. A reader would care because real structural monitoring data is almost always confounded by varying loads and weather that can mimic or mask damage effects.

Core claim

By imposing self-supervised invariance regularization via VICReg on one latent representation using baseline data where structural damage is constant but operational and environmental conditions vary, together with a frequency-domain constraint that enforces agreement between the power spectral density of the reconstructed signal and the input time series, the framework produces a disentangled representation that is sensitive to damage-related characteristics while remaining invariant to nuisance variability, enabling robust output-only damage identification and quantification.

What carries the argument

Autoencoder with two latent representations, one of which receives VICReg invariance regularization on baseline data and is further constrained by power spectral density matching to the input signal.

If this is right

  • Damage detection and quantification become possible from output-only signals without any labeled damage examples or prior knowledge of excitations.
  • The approach demonstrates robustness to operational variability on real bridge and gearbox datasets.
  • Training occurs fully end-to-end and label-free, suiting direct application to field-collected data.
  • Generalization across different structural types is observed without task-specific retraining.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same invariance-plus-spectral-constraint pattern might separate other causal factors in unlabeled sensor time series beyond structural damage.
  • If the disentanglement holds, periodic collection of new baseline data under changing conditions could maintain performance without supervised updates.
  • Deployment could reduce reliance on physics-based finite-element models for damage localization in variable environments.

Load-bearing premise

Baseline data keeps structural damage fixed while only operational and environmental conditions change, and the VICReg plus PSD constraint will produce true disentanglement without labels or further supervision.

What would settle it

Collect vibration data from a structure with known controlled damage introduction under deliberately varied excitation or temperature; the learned representation should change measurably with the damage state while remaining stable under the non-damage variations alone.

Figures

Figures reproduced from arXiv: 2604.19658 by Charikleia Stoura, Eleni Chatzi, Simon Scandella, Xudong Jian.

Figure 1
Figure 1. Figure 1: Proposed self-supervised disentangled representation learning architecture. A batch of input vibration signals X (batch size: B) is encoded into a damage-sensitive latent representation zdmg and a nuisance latent representation zndmg. The model is trained with three objectives: (1) time-domain reconstruction of X, (2) frequency-domain reconstruction to match the PSD, and (3) self-supervised invariance regu… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the openLAB bridge experiment. The top panel shows the full-scale bridge, while the middle panel illustrates the structural layout, including spans, shaker locations (S1–S2), accelerometer positions (A1–A6), the position of applied static load, and damage locations (D1–D3). The bottom panels present the cross sections of spans 1–2 (T-girder) and span 3 (hollow slab). By activating the shakers i… view at source ↗
Figure 3
Figure 3. Figure 3: Training and validation loss curves of the proposed framework on openLAB dataset. The total loss and individual loss components, including time-domain reconstruction (Loss 1), frequency-domain reconstruction (Loss 2), and VICReg self-supervised loss (Loss 3), are shown over 500 training epochs. The above settings are determined through preliminary hyperparameter tuning. It should be emphasized that the pri… view at source ↗
Figure 4
Figure 4. Figure 4: Example reconstruction results for a signal window (damage label 1, excitation label 4). Only the first 200 data points are presented to allow clearer observation: (a) Time-domain acceleration signals and their reconstructions across 12 channels. (b) Corresponding PSD and their reconstructions [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Mahalanobis distances over time for all signal windows in the validation and test sets. Colours indicate different damage states. The dashed line denotes the baseline median, and the solid line denotes the 95th percentile threshold derived from the baseline subset: (a) Results from zdmg; (b) Results from zndmg Balanced Accuracy = 1 2 (TPR + TNR) (15) [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: UMAP visualization of the learned latent representations: (a) zdmg coloured by damage state; (b) zdmg coloured by excitation type; (c) zndmg coloured by damage state; (d) zndmg coloured by excitation type. for each sensor channel, nine features are computed from the raw acceleration time series, including mean, standard deviation, skewness, kurtosis, maximum value, normalized peak frequency index, spectral… view at source ↗
Figure 7
Figure 7. Figure 7: Mahalanobis distances computed from zdmg for signal windows under excitation type 4. Colours indicate different damage states. The dashed line denotes the baseline median, and the solid line denotes the 95th percentile threshold derived from the baseline subset. Real-world Evaluation 2: MCC5 gearbox dataset This section presents the evaluation of the proposed struc￾tural damage identification framework usi… view at source ↗
Figure 8
Figure 8. Figure 8: Overview of the MCC5-THU gearbox experiment 33: (a) Experimental setup of the MCC5-THU gearbox test rig, including the motor, torque sensor, parallel gearbox, magnetic powder brake, vibration sensors, and data acquisition system. (b) Schematic of the internal gearbox structure, showing the gear arrangement and the locations of artificially introduced gear and bearing faults. sequence lasts 60 seconds and c… view at source ↗
Figure 9
Figure 9. Figure 9: Time-varying rotational speed profiles in the MCC5-THU dataset 33. Each excitation sequence lasts 60 seconds and consists of piecewise speed levels with ramp transitions. Different colours indicate different experimental groups corresponding to distinct nominal rotational speeds. Across different experiments, the predefined peak rota￾tional speeds are set to 1000 rpm, 2000 rpm, and 3000 rpm. During all exp… view at source ↗
Figure 10
Figure 10. Figure 10: presents the training and validation loss curves from a representative run, with a total training time of 161.02 seconds. As observed, all loss components decrease smoothly during training, and the validation losses remain slightly higher than the corresponding training losses, indicating stable convergence and no evident overfitting [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Example reconstruction results for a signal window (damage label 3, excitation label 1). Only the first 200 data points are presented to allow clearer observation: (a) Time-domain acceleration signals and their reconstructions across 12 channels. (b) Corresponding PSD and their reconstructions [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Mahalanobis distances for all signal windows in the validation and test sets. Colours indicate different damage states. The dashed line denotes the baseline median, and the solid line denotes the 95th percentile threshold derived from the baseline subset: (a) Results from zdmg; (b) Results from zndmg indoor conditions, resulting in reduced noise and fewer environmental variations (e.g., temperature fluctu… view at source ↗
read the original abstract

Damage identification is a core task in structural health monitoring. In practice, however, its reliability is often compromised by confounding non-damage effects, such as variations in excitation and environmental conditions, which can induce changes comparable to or larger than those caused by structural damage. To address this challenge, this study proposes a self-supervised label-free disentangled representation learning framework for robust vibration-based structural damage identification. The proposed framework employs an autoencoder with two latent representations to learn directly from raw vibration acceleration signals. A self-supervised invariance regularization, implemented via Variance-Invariance-Covariance Regularization (VICReg), is imposed on one latent representation using baseline data where structural damage is assumed constant but operational and environmental conditions vary. In addition, a frequency-domain constraint is introduced to enforce agreement between the power spectral density reconstructed from the latent representation and that computed from the corresponding input time series. Together, these mechanisms promote disentanglement, enabling the learned representation to be sensitive to damage-related characteristics while remaining invariant to nuisance variability. The framework is trained in a fully end-to-end and label-free manner, requiring no prior information on damage, excitation, or environmental conditions, making it well-suited for real-world applications. Its effectiveness is validated on two distinct real-world vibration datasets, including a bridge and a gearbox. The results demonstrate robustness to operational variability, strong generalization capability, and good performance in both damage detection and quantification.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes a label-free self-supervised disentangled representation learning framework for vibration-based structural damage identification. An autoencoder learns two latent representations from raw acceleration signals: VICReg invariance regularization is applied to one latent using baseline data (assumed to have constant/zero damage but varying operational and environmental conditions), while a power spectral density (PSD) reconstruction constraint is imposed to promote damage sensitivity in the other latent. The framework is trained end-to-end without labels or prior information and is evaluated on two real-world datasets (bridge and gearbox), with claims of robustness to operational variability, strong generalization, and good performance in damage detection and quantification.

Significance. If the disentanglement mechanism proves effective and the baseline assumption holds, the work could meaningfully advance structural health monitoring by offering a practical label-free approach to isolate damage effects from confounding variability, reducing reliance on supervised data collection in field settings.

major comments (2)
  1. [§3] §3 (framework description): The disentanglement claim rests on applying VICReg invariance regularization exclusively to one latent using baseline data asserted to contain strictly constant (zero) damage. If this premise is even partially violated, the invariance objective will treat damage-induced variations as nuisance factors and suppress them in the damage-sensitive latent; the PSD constraint provides no additional mechanism to recover the distinction. No verification procedure, sensitivity analysis, or robustness test for the baseline assumption is described, making this a load-bearing risk for the central claim.
  2. [§4] §4 (experimental validation): The abstract and results assert 'robustness,' 'strong generalization,' and 'good performance' on the bridge and gearbox datasets, yet no quantitative metrics (e.g., detection accuracy, quantification error, ROC-AUC), baseline comparisons, ablation studies isolating VICReg versus the PSD term, or error analysis are referenced. Without these, the empirical support for successful disentanglement cannot be assessed.
minor comments (2)
  1. [§3] The two latent representations are described as 'damage-sensitive' and 'invariant' but their dimensionalities, separation mechanism within the autoencoder, and exact loss weighting are not tabulated or diagrammed, which would aid reproducibility.
  2. [§4] A short table summarizing the two datasets (sampling rates, number of recordings, known damage states) would improve clarity in the experimental section.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's constructive feedback on our manuscript. We address each major comment below with planned revisions to strengthen the presentation of the framework and its validation.

read point-by-point responses

  1. Referee: [§3] §3 (framework description): The disentanglement claim rests on applying VICReg invariance regularization exclusively to one latent using baseline data asserted to contain strictly constant (zero) damage. If this premise is even partially violated, the invariance objective will treat damage-induced variations as nuisance factors and suppress them in the damage-sensitive latent; the PSD constraint provides no additional mechanism to recover the distinction. No verification procedure, sensitivity analysis, or robustness test for the baseline assumption is described, making this a load-bearing risk for the central claim.

    Authors: We agree that the baseline assumption is central to the disentanglement mechanism. In the bridge and gearbox datasets, baseline recordings are taken from structures confirmed to be undamaged at the time of collection, with variability arising solely from operational and environmental factors. To directly address the concern, we will add a sensitivity analysis in the revised experimental section. This will involve introducing controlled levels of simulated damage into baseline samples and quantifying the resulting impact on the separation between the invariant and damage-sensitive latents, thereby testing robustness to partial violations of the assumption. revision: partial

  2. Referee: [§4] §4 (experimental validation): The abstract and results assert 'robustness,' 'strong generalization,' and 'good performance' on the bridge and gearbox datasets, yet no quantitative metrics (e.g., detection accuracy, quantification error, ROC-AUC), baseline comparisons, ablation studies isolating VICReg versus the PSD term, or error analysis are referenced. Without these, the empirical support for successful disentanglement cannot be assessed.

    Authors: The current manuscript presents results primarily through visualizations of damage detection and quantification performance across the two datasets. We acknowledge that explicit numerical metrics, comparisons, and ablations would strengthen the empirical claims. In the revision, we will add tables reporting detection accuracy, mean quantification error, and ROC-AUC values. We will also include comparisons against standard autoencoder baselines and traditional signal-processing methods, plus ablation experiments that isolate the VICReg invariance term from the PSD reconstruction constraint. A dedicated error analysis subsection will discuss observed failure modes and their relation to operational variability. revision: yes

Circularity Check

0 steps flagged

No circularity: framework is an independent proposal with external validation

full rationale

The paper presents a self-supervised autoencoder framework using VICReg invariance on baseline data (assumed constant damage) plus a PSD reconstruction constraint to promote disentanglement. No equations, derivations, or self-citations are shown that reduce the claimed sensitivity/invariance properties to a fitted parameter, renamed input, or self-referential definition by construction. Training occurs end-to-end on external real-world datasets (bridge, gearbox) without the target outputs being presupposed in the inputs. The baseline-damage assumption is a modeling premise subject to empirical verification, not a circular reduction. This qualifies as a standard non-circular proposal of a new method.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Ledger is inferred from the abstract alone; the central claim rests on domain assumptions about baseline data properties and the disentangling power of the chosen regularizers rather than on new free parameters or invented entities.

axioms (2)
  • domain assumption Baseline data has constant structural damage while operational and environmental conditions vary
    Invoked to apply VICReg invariance regularization that isolates damage effects
  • domain assumption VICReg regularization plus PSD frequency constraint will produce damage-sensitive yet variability-invariant latent representations
    Core premise that the two mechanisms together achieve the desired disentanglement without labels

pith-pipeline@v0.9.0 · 5570 in / 1454 out tokens · 42761 ms · 2026-05-10T03:44:42.771013+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages

  1. [1]

    Recent progress and future trends on damage identification methods for bridge structures.Structural Control and Health Monitoring2019; 26(10): e2416

    An Y , Chatzi E, Sim SH et al. Recent progress and future trends on damage identification methods for bridge structures.Structural Control and Health Monitoring2019; 26(10): e2416

    work page

  2. [2]

    An introduction to structural health monitoring.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences2007; 365(1851): 303–315

    Farrar CR and Worden K. An introduction to structural health monitoring.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences2007; 365(1851): 303–315

    work page

  3. [3]

    Vibration based condition monitoring: a review.Structural health monitoring2004; 3(4): 355–377

    Carden EP and Fanning P. Vibration based condition monitoring: a review.Structural health monitoring2004; 3(4): 355–377

    work page

  4. [4]

    Doebling SW, Farrar CR, Prime MB et al. Damage identifi- cation and health monitoring of structural and mechanical sys- tems from changes in their vibration characteristics: a literature review.Los Alamos National Laboratory Report LA-13070-MS 1996

    work page 1996

  5. [5]

    John Wiley & Sons, 2012

    Farrar CR and Worden K.Structural health monitoring: a machine learning perspective. John Wiley & Sons, 2012

    work page 2012

  6. [6]

    Sohn H. Effects of environmental and operational variability on structural health monitoring.Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences2007; 365(1851): 539–560

    work page

  7. [7]

    The time-varying modal information of a cable-stayed bridge: some consideration for shm.Engineering Structures2021; 235: 111835

    Li Y , Zhu L, Qian C et al. The time-varying modal information of a cable-stayed bridge: some consideration for shm.Engineering Structures2021; 235: 111835

    work page

  8. [8]

    Machine learning algorithms for damage detection under operational and environmental variability.Structural Health Monitoring2011; 10(6): 559–572

    Figueiredo E, Park G, Farrar CR et al. Machine learning algorithms for damage detection under operational and environmental variability.Structural Health Monitoring2011; 10(6): 559–572

    work page

  9. [9]

    One-year monitoring of the z24-bridge: environmental effects versus damage events

    Peeters B and De Roeck G. One-year monitoring of the z24-bridge: environmental effects versus damage events. Earthquake engineering & structural dynamics2001; 30(2): 149–171

    work page

  10. [10]

    Output- only structural health monitoring in changing environmental conditions by means of nonlinear system identification

    Reynders E, Wursten G and De Roeck G. Output- only structural health monitoring in changing environmental conditions by means of nonlinear system identification. Structural Health Monitoring2014; 13(1): 82–93

    work page

  11. [11]

    Vibration- based structural health monitoring using output-only measure- ments under changing environment.Mechanical systems and signal processing2008; 22(1): 34–56

    Deraemaeker A, Reynders E, De Roeck G et al. Vibration- based structural health monitoring using output-only measure- ments under changing environment.Mechanical systems and signal processing2008; 22(1): 34–56

    work page

  12. [12]

    Cross EJ, Worden K and Chen Q. Cointegration: a novel approach for the removal of environmental trends in structural health monitoring data.Proceedings of the Royal Society A: Mathematical, Physical and Engineering Sciences2011; 467(2133): 2712–2732

    work page

  13. [13]

    Shi H, Worden K and Cross EJ. A regime-switching cointe- gration approach for removing environmental and operational variations in structural health monitoring.Mechanical Systems and Signal Processing2018; 103: 381–397

    work page

  14. [14]

    Unsupervised learning methods for data-driven vibration-based structural health monitoring: a review.Sensors2023; 23(6): 3290

    Eltouny K, Gomaa M and Liang X. Unsupervised learning methods for data-driven vibration-based structural health monitoring: a review.Sensors2023; 23(6): 3290

    work page

  15. [15]

    Unsupervised anomaly detection based on deep autoencoders, information fusion, and active sensing.Structural Health Monitoring2026; : 14759217251410304

    Wang Z, Khokhar SA, Jahanshahi MR et al. Unsupervised anomaly detection based on deep autoencoders, information fusion, and active sensing.Structural Health Monitoring2026; : 14759217251410304

    work page

  16. [16]

    Cadini F, Lomazzi L, Roca MF et al. Neutralization of temperature effects in damage diagnosis of mdof systems by combinations of autoencoders and particle filters.Mechanical Systems and Signal Processing2022; 162: 108048

    work page

  17. [17]

    A domain adaptation approach to damage classification with an application to bridge monitoring.Mechanical Systems and Signal Processing2024; 209: 111135

    Giglioni V , Poole J, Venanzi I et al. A domain adaptation approach to damage classification with an application to bridge monitoring.Mechanical Systems and Signal Processing2024; 209: 111135

    work page

  18. [18]

    Fusing damage- sensitive features and domain adaptation towards robust damage classification in real buildings.Soil Dynamics and Earthquake Engineering2023; 166: 107739

    Martakis P, Reuland Y , Stavridis A et al. Fusing damage- sensitive features and domain adaptation towards robust damage classification in real buildings.Soil Dynamics and Earthquake Engineering2023; 166: 107739

    work page

  19. [19]

    Structural damage classification under varying environmental conditions and unknown classes via open set domain adaptation.Mechanical Systems and Signal Processing2024; 218: 111561

    Zhou M and Lai Z. Structural damage classification under varying environmental conditions and unknown classes via open set domain adaptation.Mechanical Systems and Signal Processing2024; 218: 111561

    work page

  20. [20]

    Towards unsupervised fault detection for offshore wind turbine cable protection systems using contrastive learning

    Duth ´e G, L’Homme Y , Abdallah I et al. Towards unsupervised fault detection for offshore wind turbine cable protection systems using contrastive learning. InJournal of Physics: Conference Series, volume 2767. IOP Publishing, p. 032038

    work page

  21. [21]

    Transferring self-supervised pretrained models for shm data anomaly detection with scarce labeled data.Structural Control and Health Monitoring2025; 2025(1): 2414195

    Zhou M, Jian X, Xia Y et al. Transferring self-supervised pretrained models for shm data anomaly detection with scarce labeled data.Structural Control and Health Monitoring2025; 2025(1): 2414195

    work page 2025

  22. [22]

    Contrastive and self-supervised learning for open-set damage classification in structural health monitoring with incomplete and imbalanced vibration data

    Shi S, Du D, Mercan O et al. Contrastive and self-supervised learning for open-set damage classification in structural health monitoring with incomplete and imbalanced vibration data. Expert Systems with Applications2025; 293: 128731

    work page

  23. [23]

    Domain-separated capsule network for damage detection in aluminum plates under varying vibration conditions.Ultrasonics2025; 154: 107688

    Jiang Q, Huang X, Qu W et al. Domain-separated capsule network for damage detection in aluminum plates under varying vibration conditions.Ultrasonics2025; 154: 107688

    work page

  24. [24]

    Disentangled representation learning with physics-informed variational autoencoder for structural health monitoring

    Koune I and Cicirello A. Disentangled representation learning with physics-informed variational autoencoder for structural health monitoring. InProceedings of the 11th European Workshop on Structural Health Monitoring (EWSHM 2024)

    work page 2024

  25. [25]

    Bel-Hadj Y , Weijtjens W and Devriendt C. Structural health monitoring in a population of similar structures with self-supervised learning: a two-stage approach for enhanced damage detection and model tuning.Structural Health Monitoring2025; : 14759217251324194

    work page

  26. [26]

    Vicreg: Variance-invariance- covariance regularization for self-supervised learning

    Bardes A, Ponce J and Lecun Y . Vicreg: Variance-invariance- covariance regularization for self-supervised learning. InICLR 2022-International Conference on Learning Representations

    work page 2022

  27. [27]

    Umap: Uniform manifold approximation and projection for dimension reduction.arXiv Prepared usingsagej.cls 18 Journal Title XX(X) preprint arXiv:1802034262018

    McInnes L, Healy J and Melville J. Umap: Uniform manifold approximation and projection for dimension reduction.arXiv Prepared usingsagej.cls 18 Journal Title XX(X) preprint arXiv:1802034262018

    work page

  28. [28]

    A comparison of linear approaches to filter out environmental effects in structural health monitoring.Mechanical systems and signal processing 2018; 105: 1–15

    Deraemaeker A and Worden K. A comparison of linear approaches to filter out environmental effects in structural health monitoring.Mechanical systems and signal processing 2018; 105: 1–15

    work page 2018

  29. [29]

    openlab– eine forschungsbr ¨ucke zur entwicklung eines digitalen br¨uckenzwillings.Beton-und Stahlbetonbau2024; 119(3): 169–180

    Herbers M, Bartels JH, Richter B et al. openlab– eine forschungsbr ¨ucke zur entwicklung eines digitalen br¨uckenzwillings.Beton-und Stahlbetonbau2024; 119(3): 169–180

    work page

  30. [30]

    openlab bridge bautzen germany - load test 05th - 07th may 2025 - vibration excitation measurements, 2026

    Herrmann R, Hille F and Degener S. openlab bridge bautzen germany - load test 05th - 07th may 2025 - vibration excitation measurements, 2026. DOI:10.5281/zenodo.18546379. URL https://doi.org/10.5281/zenodo.18546379

    work page doi:10.5281/zenodo.18546379 2025

  31. [31]

    Ssrl: Self-supervised representation learning for structural damage identification.https://github.com/ JxdEngineer/SSRL, 2025

    Jian X. Ssrl: Self-supervised representation learning for structural damage identification.https://github.com/ JxdEngineer/SSRL, 2025. GitHub repository, accessed 2026

    work page 2025

  32. [32]

    Visualizing data using t-sne

    Van der Maaten L and Hinton G. Visualizing data using t-sne. Journal of machine learning research2008; 9(11)

    work page

  33. [33]

    Multi-mode fault diagnosis datasets of gearbox under variable working conditions.Data in brief 2024; 54: 110453

    Chen S, Liu Z, He X et al. Multi-mode fault diagnosis datasets of gearbox under variable working conditions.Data in brief 2024; 54: 110453. Prepared usingsagej.cls

    work page 2024