pith. sign in

arxiv: 2507.12969 · v2 · pith:CYQEGY52new · submitted 2025-07-17 · 💻 cs.LG · cs.CV

WaveletInception Networks for on-board Vibration-Based Infrastructure Health Monitoring

Reza Riahi Samani , Alfredo Nunez , Bart De Schutter This is my paper

Pith reviewed 2026-05-21 23:52 UTC · model grok-4.3

classification 💻 cs.LG cs.CV
keywords vibration signal analysisinfrastructure health monitoringlearnable wavelet transformdeep learningBiGRUon-board monitoringtrack stiffnesstransition zone classification
0
0 comments X

The pith

A learnable wavelet network processes raw on-board vibration signals at varying speeds to produce high-resolution infrastructure health maps without explicit preprocessing.

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

The paper develops a WaveletInception-BiGRU architecture that extracts spectral features through a learnable wavelet packet transform, learns multi-scale representations with one-dimensional Inception-ResNet blocks, and folds in operational variables such as measurement speed inside bidirectional GRU layers. This pipeline turns variable-speed vibration recordings into localized health estimates and spatially aligned profiles for infrastructure elements. A sympathetic reader would care because the method removes the usual requirement for speed normalization or filtering steps, allowing continuous automated monitoring from moving vehicles. Case studies on real rail data show improved regression of track stiffness and classification of transition zones compared with prior approaches.

Core claim

The WaveletInception-BiGRU network performs early spectral feature extraction with a Learnable Wavelet Packet Transform, followed by 1D Inception-ResNet modules for multi-scale feature learning and BiGRU modules that integrate temporal dependencies while incorporating operational conditions such as measurement speed. This sequence enables direct analysis of vibration signals recorded at varying speeds without explicit preprocessing and produces accurate localized health assessments together with high-resolution profiles spatially mapped to the physical infrastructure layout.

What carries the argument

WaveletInception-BiGRU network that uses a learnable wavelet packet transform for initial spectral features, Inception-ResNet blocks for multi-scale extraction, and speed-aware BiGRU layers for temporal integration before a sequential estimation head.

If this is right

  • Vibration data collected at different vehicle speeds can be used directly for health assessment without separate normalization steps.
  • High-resolution health profiles can be generated that align with the physical geometry of the monitored structure.
  • The same pipeline supports both regression tasks such as track stiffness estimation and classification tasks such as transition-zone detection.
  • On-board monitoring becomes feasible because the network operates on raw signals and accounts for operational variables internally.

Where Pith is reading between the lines

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

  • The approach may extend to other linear infrastructures such as pipelines or overhead lines if similar vibration patterns are available.
  • Real-time deployment on inspection vehicles could reduce the need for periodic manual surveys by providing continuous spatial health maps.
  • Integration with existing sensor networks on trains or maintenance cars would allow the method to operate without dedicated data-collection campaigns.

Load-bearing premise

That embedding measurement speed inside the BiGRU layers along with learnable wavelet features is enough to deliver accurate localized health estimates on new infrastructures and unseen operating conditions without extra tuning or filtering.

What would settle it

Performance drop on vibration recordings from a different infrastructure type or from speeds far outside the training distribution while keeping the same network architecture and training procedure.

Figures

Figures reproduced from arXiv: 2507.12969 by Alfredo Nunez, Bart De Schutter, Reza Riahi Samani.

Figure 1
Figure 1. Figure 1: Tree-like structure of the Wavelet Packet Transform (WPT). [PITH_FULL_IMAGE:figures/full_fig_p011_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Long Short-Term Memory (LSTM) cell. The fundamental components of an [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: 1D Inception block structure. 4.2.2. Feature Fusion and Temporal Features To incorporate the operational conditions of vibration signal measure￾ment, we propose using an LSTM layer that integrates features from different sources. The operational conditions, such as measurement speed, are repre￾sented in a feature vector through an embedding layer. This embedding is then concatenated with the extracted vibr… view at source ↗
Figure 4
Figure 4. Figure 4: Overview of the proposed methodology, including WaveletInception vibration [PITH_FULL_IMAGE:figures/full_fig_p018_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: The layout of the 10 sleepers track segment. [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Representation of the simulated ABA signals in four different measurement [PITH_FULL_IMAGE:figures/full_fig_p023_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Average ABA signal power in the scenario of stiffness reduction in three sleepers: [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Loss convergence on the validation set for the proposed model and baselines. [PITH_FULL_IMAGE:figures/full_fig_p026_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: WI-BiLSTM model’s predictions on the test set, in the scenarios of constant [PITH_FULL_IMAGE:figures/full_fig_p029_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Distribution of learned DWT coefficients. The top figure shows the distribution [PITH_FULL_IMAGE:figures/full_fig_p032_10.png] view at source ↗
read the original abstract

This paper presents a deep learning framework for analyzing on board vibration response signals in infrastructure health monitoring. The proposed WaveletInception-BiGRU network uses a Learnable Wavelet Packet Transform (LWPT) for early spectral feature extraction, followed by one-dimensional Inception-Residual Network (1D Inception-ResNet) modules for multi-scale, high-level feature learning. Bidirectional Gated Recurrent Unit (BiGRU) modules then integrate temporal dependencies and incorporate operational conditions, such as the measurement speed. This approach enables effective analysis of vibration signals recorded at varying speeds, eliminating the need for explicit signal preprocessing. The sequential estimation head further leverages bidirectional temporal information to produce an accurate, localized assessment of infrastructure health. Ultimately, the framework generates high-resolution health profiles spatially mapped to the physical layout of the infrastructure. Case studies involving track stiffness regression and transition zone classification using real-world measurements demonstrate that the proposed framework significantly outperforms state-of-the-art methods, underscoring its potential for accurate, localized, and automated on-board infrastructure health monitoring.

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 presents a deep learning framework called WaveletInception-BiGRU for on-board vibration-based infrastructure health monitoring. It combines a Learnable Wavelet Packet Transform (LWPT) for early spectral feature extraction, 1D Inception-ResNet modules for multi-scale feature learning, and BiGRU modules that integrate temporal dependencies and operational conditions such as measurement speed. The approach claims to handle vibration signals at varying speeds without explicit preprocessing and to produce high-resolution, spatially mapped health profiles. Case studies on real-world track stiffness regression and transition zone classification demonstrate significant outperformance over state-of-the-art methods.

Significance. If the empirical results hold under rigorous validation, the work could meaningfully advance automated on-board monitoring by reducing preprocessing requirements and directly conditioning on operational variables. The combination of learnable wavelets with speed-aware recurrent modules represents a targeted technical contribution for vibration analysis under variable conditions. No mention is made of open code, reproducible pipelines, or parameter-free derivations, which limits the assessed impact relative to papers that provide such assets.

major comments (2)
  1. [§4 (Case Studies) and §3.3 (BiGRU Integration)] The central claim that LWPT features plus BiGRU conditioning on measurement speed suffice to produce accurate localized health maps from raw varying-speed signals (without preprocessing) is load-bearing but under-supported. No evidence is presented of speed-stratified cross-validation, ablation removing the speed input, or testing on infrastructure/speed ranges outside the training distribution. Physical scaling of vibration amplitude and frequency with speed (via wheel-rail dynamics) creates a risk that the model exploits spurious speed-health correlations rather than invariant features.
  2. [Abstract and §4] The abstract and results sections assert significant outperformance on track stiffness regression and transition classification, yet the manuscript provides no quantitative metrics, error bars, baseline implementation details, or description of data splits and train/test partitioning. This prevents verification of the performance claims against the paper's own evidence and directly affects the soundness of the empirical contribution.
minor comments (2)
  1. [§3.1] Notation for the LWPT is introduced without a clear equation defining the learnable parameters or the wavelet packet decomposition tree; a single equation or diagram would improve clarity.
  2. [Figure 2 and Figure 4] Figure captions for the network architecture and health-profile outputs should explicitly state the input dimensions, output resolution, and any spatial mapping procedure used to align predictions with infrastructure layout.

Simulated Author's Rebuttal

2 responses · 1 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. These have prompted us to strengthen the empirical validation and presentation of results. We address each major comment below and indicate the revisions planned for the next version.

read point-by-point responses

  1. Referee: [§4 (Case Studies) and §3.3 (BiGRU Integration)] The central claim that LWPT features plus BiGRU conditioning on measurement speed suffice to produce accurate localized health maps from raw varying-speed signals (without preprocessing) is load-bearing but under-supported. No evidence is presented of speed-stratified cross-validation, ablation removing the speed input, or testing on infrastructure/speed ranges outside the training distribution. Physical scaling of vibration amplitude and frequency with speed (via wheel-rail dynamics) creates a risk that the model exploits spurious speed-health correlations rather than invariant features.

    Authors: We agree that additional targeted validation is needed to substantiate the claim and mitigate concerns about spurious correlations. In the revised manuscript we will add (i) an ablation study that removes the speed input from the BiGRU while keeping all other components fixed, (ii) speed-stratified cross-validation results on the existing datasets to demonstrate consistent performance across speed bins, and (iii) an expanded discussion of the physical scaling effects together with evidence that the learned features remain informative after speed conditioning. We acknowledge that testing on entirely new infrastructure types and speed ranges outside the current collection is not feasible with the data at hand; we will therefore add a limitations paragraph highlighting this generalization gap and the value of future multi-site validation. revision: partial

  2. Referee: [Abstract and §4] The abstract and results sections assert significant outperformance on track stiffness regression and transition classification, yet the manuscript provides no quantitative metrics, error bars, baseline implementation details, or description of data splits and train/test partitioning. This prevents verification of the performance claims against the paper's own evidence and directly affects the soundness of the empirical contribution.

    Authors: We apologize for the insufficient detail in the abstract and the compressed presentation of experimental settings. The full results in Section 4 contain the underlying numbers, but we will revise the abstract to report the key quantitative metrics (e.g., RMSE and classification accuracy with standard deviations across folds). In the revised Section 4 we will also provide explicit descriptions of baseline re-implementations (following the original authors' protocols), the precise train/test partitioning strategy, and the cross-validation procedure used to generate error bars. These additions will make the performance claims directly verifiable from the manuscript. revision: yes

standing simulated objections not resolved
  • Testing on infrastructure and speed ranges completely outside the current training distribution cannot be performed because additional real-world datasets from new sites are not available to the authors at this time.

Circularity Check

0 steps flagged

No circularity in derivation chain

full rationale

The paper presents an empirical deep learning architecture (LWPT + 1D Inception-ResNet + BiGRU with speed conditioning) evaluated on real-world track stiffness and transition classification tasks. No equations, uniqueness theorems, or self-citations are invoked to derive performance claims; results are reported from direct training and testing on measured data without reduction to fitted inputs or prior author work by construction. The framework is self-contained as a data-driven model whose validity rests on external benchmarks rather than internal redefinition.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 1 invented entities

The central claim rests on the untested assumption that the described neural architecture generalizes from the two mentioned case studies to broader infrastructure without post-hoc adjustments. No explicit free parameters, axioms, or invented entities are detailed in the abstract.

invented entities (1)
  • Learnable Wavelet Packet Transform (LWPT) no independent evidence
    purpose: Early spectral feature extraction from raw vibration signals
    Introduced as a core component of the framework; no independent evidence of its properties outside the paper is provided in the abstract.

pith-pipeline@v0.9.0 · 5714 in / 1338 out tokens · 38276 ms · 2026-05-21T23:52:15.885446+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    The WaveletInception feature extractor utilizes a Learnable Wavelet Packet Transform (LWPT) as the stem for extracting vibration signal features, followed by 1D Inception networks... BiLSTM... integrate operational conditions such as the measurement speed.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

51 extracted references · 51 canonical work pages · 4 internal anchors

  1. [1]

    C. from the Commission to the European Parliament, the Council, Eighth monitoring report on the development of the rail market un- der Article 15(4) of Directive 2012/34/EU of the European Parliament and of the Council (2023)

    work page 2012

  2. [2]

    H. A. Fern´ andez-Bobadilla, U. Martin, Modern tendencies in vehicle- based condition monitoring of the railway track, IEEE Trans- actions on Instrumentation and Measurement 72 (2023) 1–44. doi:10.1109/TIM.2023.3243673

    work page doi:10.1109/tim.2023.3243673 2023

  3. [3]

    Sansi˜ nena, B

    A. Sansi˜ nena, B. Rodr´ ıguez-Arana, S. Arrizabalaga, A systematic review of acceleration-based estimation of railway track quality, Vehicle System Dynamics (2025) 1–28doi:10.1080/00423114.2025.2483972

    work page doi:10.1080/00423114.2025.2483972 2025

  4. [4]

    Phusakulkajorn, S

    W. Phusakulkajorn, S. Unsiwilai, L. Chang, Z. Li, A. N´ u˜ nez, A hybrid neural model approach for health assessment of railway transition zones with multiple data sources, IEEE Transactions on Instrumentation and Measurement 74 (2025) 1–17. doi:10.1109/TIM.2025.3575986

    work page doi:10.1109/tim.2025.3575986 2025

  5. [5]

    Phusakulkajorn, A

    W. Phusakulkajorn, A. N´ u˜ nez, H. Wang, A. Jamshidi, A. Zoeteman, B. Ripke, R. Dollevoet, B. De Schutter, Z. Li, Artificial intelligence in railway infrastructure: Current research, challenges, and future op- portunities, Intelligent Transportation Infrastructure 2 (2023) liad016. doi:10.1093/iti/liad016

    work page doi:10.1093/iti/liad016 2023

  6. [6]

    Lederman, S

    G. Lederman, S. Chen, J. Garrett, J. Kovaˇ cevi´ c, H. Y. Noh, J. Bielak, Track-monitoring from the dynamic response of an operational train 87 (2017-03) 1–16. doi:10.1016/j.ymssp.2016.06.041

    work page doi:10.1016/j.ymssp.2016.06.041 2017

  7. [7]

    C. Shen, P. Zhang, R. Dollevoet, A. Zoeteman, Z. Li, Evaluating rail- way track stiffness using axle box accelerations: A digital twin ap- proach, Mechanical Systems and Signal Processing 204 (2023) 110730. doi:10.1016/j.ymssp.2023.110730

    work page doi:10.1016/j.ymssp.2023.110730 2023

  8. [8]

    A. C. Lamprea-Pineda, D. P. Connolly, A. Castanheira-Pinto, P. Alves- Costa, M. F. Hussein, P. K. Woodward, On railway track receptance 177 (2024-02) 108331. doi:10.1016/j.soildyn.2023.108331. 36

    work page doi:10.1016/j.soildyn.2023.108331 2024

  9. [9]

    Unsiwilai, L

    S. Unsiwilai, L. Wang, A. N´ u˜ nez, Z. Li, Multiple-axle box accelera- tion measurements at railway transition zones 213 (2023-05) 112688. doi:10.1016/j.measurement.2023.112688

    work page doi:10.1016/j.measurement.2023.112688 2023

  10. [10]

    Y. B. Yang, J. P. Yang, State-of-the-Art Review on Modal Identification and Damage Detection of Bridges by Moving Test Vehicles 18 (2018-02) 1850025. doi:10.1142/S0219455418500256

    work page doi:10.1142/s0219455418500256 2018

  11. [11]

    Quirke, D

    P. Quirke, D. Cantero, E. J. OBrien, C. Bowe, Drive-by detection of railway track stiffness variation using in-service vehicles, Proceedings of the Institution of Mechanical Engineers, Part F: Journal of Rail and Rapid Transit 231 (2017) 498–514. doi:10.1177/0954409716634752

    work page doi:10.1177/0954409716634752 2017

  12. [12]

    X. Q. Zhu, S. S. Law, L. Huang, Identification of Railway Ballasted Track Systems from Dynamic Responses of In-Service Trains, Journal of Aerospace Engineering 31 (2018) 04018060. doi:10.1061/(ASCE)AS.1943-5525.0000898

    work page doi:10.1061/(asce)as.1943-5525.0000898 2018

  13. [13]

    Caprioli, A

    A. Caprioli, A. Cigada, D. Raveglia, Rail inspection in track maintenance: A benchmark between the wavelet approach and the more conventional Fourier analysis 21 (2007-02) 631–652. doi:10.1016/j.ymssp.2005.12.001

    work page doi:10.1016/j.ymssp.2005.12.001 2007

  14. [14]

    C. Shen, R. Dollevoet, Z. Li, Fast and robust identification of railway track stiffness from simple field measurement, Mechanical Systems and Signal Processing 152 (2021) 107431

    work page 2021

  15. [15]

    Hoelzl, V

    C. Hoelzl, V. Dertimanis, L. Ancu, A. Kollros, E. Chatzi, Vold–Kalman filter order tracking of axle box accelerations for track stiffness assess- ment, Mechanical Systems and Signal Processing 204 (2023) 110817. doi:10.1016/j.ymssp.2023.110817

    work page doi:10.1016/j.ymssp.2023.110817 2023

  16. [16]

    Malekjafarian, E

    A. Malekjafarian, E. OBrien, P. Quirke, C. Bowe, Railway Track Mon- itoring Using Train Measurements: An Experimental Case Study 9 (2019-01) 4859. doi:10.3390/app9224859

    work page doi:10.3390/app9224859 2019

  17. [17]

    X. Mao, C. Xia, J. Liu, H. Zhang, Y. Ding, Y. Yao, Z. Liu, A novel similarity measure based on dispersion-transition matrix and Jensen–Fisher divergence and its application on the detection of rail short-wave defects, Chaos, Solitons & Fractals 192 (2025) 115988. doi:10.1016/j.chaos.2025.115988. 37

    work page doi:10.1016/j.chaos.2025.115988 2025

  18. [18]

    Locke, J

    W. Locke, J. Sybrandt, L. Redmond, I. Safro, S. Atamturktur, Using drive-by health monitoring to detect bridge damage consid- ering environmental and operational effects 468 (2020-03) 115088. doi:10.1016/j.jsv.2019.115088

    work page doi:10.1016/j.jsv.2019.115088 2020

  19. [19]

    Huang, X

    J. Huang, X. Yin, S. Kaewunruen, Quantification of Dynamic Track Stiffness Using Machine Learning 10 (2022) 78747–78753. doi:10.1109/ACCESS.2022.3191278

    work page doi:10.1109/access.2022.3191278 2022

  20. [20]

    Hajializadeh, Deep learning-based indirect bridge damage identifica- tion system 22 (2023-03) 897–912

    D. Hajializadeh, Deep learning-based indirect bridge damage identifica- tion system 22 (2023-03) 897–912. doi:10.1177/14759217221087147

    work page doi:10.1177/14759217221087147 2023

  21. [21]

    Corbally, A

    R. Corbally, A. Malekjafarian, A deep-learning framework for classify- ing the type, location, and severity of bridge damage using drive-by measurements 39 (2024) 852–871. doi:10.1111/mice.13104

    work page doi:10.1111/mice.13104 2024

  22. [22]

    Phusakulkajorn, Y

    W. Phusakulkajorn, Y. Zeng, Z. Li, A. N´ u˜ nez, Unsupervised represen- tation learning for monitoring rail infrastructures with high-frequency moving vibration sensors, IEEE Transactions on Intelligent Transporta- tion Systems (2025) 1–15doi:10.1109/TITS.2025.3557712

    work page doi:10.1109/tits.2025.3557712 2025

  23. [23]

    R. R. Samani, A. Nunez, B. D. Schutter, A Bidirectional Long Short Term Memory Approach for Infrastructure Health Monitor- ing Using On-board Vibration Response (2024-12). arXiv:2412.02643, doi:10.48550/arXiv.2412.02643

    work page doi:10.48550/arxiv.2412.02643 2024

  24. [24]

    Michau, G

    G. Michau, G. Frusque, O. Fink, Fully learnable deep wavelet transform for unsupervised monitoring of high frequency time series 119(8) (2022-

    work page 2022

  25. [25]

    Frusque, O

    G. Frusque, O. Fink, Robust time series denoising with learnable wavelet packet transform 62 (2024-10) 102669. doi:10.1016/j.aei.2024.102669

    work page doi:10.1016/j.aei.2024.102669 2024

  26. [26]

    P. Liu, H. Zhang, W. Lian, W. Zuo, Multi-level wavelet convolutional neural networks 7 (2019) 74973–74985

    work page 2019

  27. [27]

    T. Li, Z. Zhao, C. Sun, L. Cheng, X. Chen, R. Yan, R. X. Gao, WaveletKernelNet: An Interpretable Deep Neural Net- work for Industrial Intelligent Diagnosis 52 (2022-04) 2302–2312. doi:10.1109/TSMC.2020.3048950. 38

    work page doi:10.1109/tsmc.2020.3048950 2022

  28. [28]

    L. E. bouny, M. Khalil, A. Adib, An End-to-End Multi-Level Wavelet Convolutional Neural Networks for heart diseases diagnosis 417 (2020-

    work page 2020

  29. [29]

    doi:10.1016/j.neucom.2020.07.056

    187–201. doi:10.1016/j.neucom.2020.07.056

    work page doi:10.1016/j.neucom.2020.07.056 2020

  30. [30]

    T. Yao, Y. Pan, Y. Li, C.-W. Ngo, T. Mei, Wave-ViT: Unifying Wavelet and Transformers for Visual Representation Learning, in: S. Avidan, G. Brostow, M. Ciss´ e, G. M. Farinella, T. Hassner (Eds.), Computer Vision – ECCV 2022, 2022, pp. 328–345. doi:10.1007/978-3-031-19806- 9 19

    work page doi:10.1007/978-3-031-19806- 2022

  31. [31]

    J. Wang, Z. Wang, J. Li, J. Wu, Multilevel Wavelet Decomposi- tion Network for Interpretable Time Series Analysis, in: Proceed- ings of the 24th ACM SIGKDD International Conference on Knowl- edge Discovery & Data Mining, KDD ’18, 2018-07, pp. 2437–2446. doi:10.1145/3219819.3220060

    work page doi:10.1145/3219819.3220060 2018

  32. [32]

    Mallat, A Wavelet Tour of Signal Processing, Elsevier, 1999

    S. Mallat, A Wavelet Tour of Signal Processing, Elsevier, 1999. doi:10.1016/B978-0-12-374370-1.X0001-8

    work page doi:10.1016/b978-0-12-374370-1.x0001-8 1999

  33. [33]

    C. C. Aggarwal, et al., Neural Networks and Deep Learning, Vol. 10, Springer, 2018

    work page 2018

  34. [34]

    Palma, Long-memory Time Series: Theory and Methods, John Wi- ley & Sons, 2007

    W. Palma, Long-memory Time Series: Theory and Methods, John Wi- ley & Sons, 2007. doi:10.1002/9780470131466

    work page doi:10.1002/9780470131466 2007

  35. [35]

    Long short-term memory,

    S. Hochreiter, J. Schmidhuber, Long Short-Term Memory 9 (1997-11) 1735–1780. doi:10.1162/neco.1997.9.8.1735

    work page doi:10.1162/neco.1997.9.8.1735 1997

  36. [36]

    Lahat, T

    D. Lahat, T. Adali, C. Jutten, Multimodal data fusion: an overview of methods, challenges, and prospects, Proceedings of the IEEE 103 (9) (2015) 1449–1477

    work page 2015

  37. [37]

    Zhang, F

    R. Zhang, F. Nie, X. Li, X. Wei, Feature selection with multi-view data: A survey 50 (2019) 158–167. doi:10.1016/j.inffus.2018.11.019

    work page doi:10.1016/j.inffus.2018.11.019 2019

  38. [38]

    H. V. Dang, H. Tran-Ngoc, T. V. Nguyen, T. Bui-Tien, G. De Roeck, H. X. Nguyen, Data-Driven Structural Health Monitoring Using Fea- ture Fusion and Hybrid Deep Learning 18 (2021-10) 2087–2103. doi:10.1109/TASE.2020.3034401. 39

    work page doi:10.1109/tase.2020.3034401 2021

  39. [39]

    L. Mou, C. Zhou, P. Zhao, B. Nakisa, M. N. Rastgoo, R. Jain, W. Gao, Driver stress detection via multimodal fusion using attention-based CNN-LSTM 173 (2021-07) 114693. doi:10.1016/j.eswa.2021.114693

    work page doi:10.1016/j.eswa.2021.114693 2021

  40. [40]

    Unsiwilai, W

    S. Unsiwilai, W. Phusakulkajorn, C. Shen, A. Zoeteman, R. Dollevoet, A. N´ u˜ nez, Z. Li, Enhanced vertical railway track quality index with dynamic responses from moving trains 10 (2024-10) e38670. doi:10.1016/j.heliyon.2024.e38670

    work page doi:10.1016/j.heliyon.2024.e38670 2024

  41. [41]

    C. Shi, Y. Zhou, L. Xu, X. Zhang, Y. Guo, A critical review on the vertical stiffness irregularity of railway ballasted track 400 (2023-10) 132715. doi:10.1016/j.conbuildmat.2023.132715

    work page doi:10.1016/j.conbuildmat.2023.132715 2023

  42. [42]

    R. R. Samani, Data underlying the publication: Waveletinception networks for drive-by vibration-based infrastructure health monitor- ing, 4tu.researchdata (2025). doi:10.4121/8d6245b5-2d81-436f-896c- 83a78e22d213

    work page doi:10.4121/8d6245b5-2d81-436f-896c- 2025

  43. [43]

    Gonz´ alez, K

    A. Gonz´ alez, K. Feng, M. Casero, Effective separation of vehicle, road and bridge information from drive-by acceleration data via the power spectral density resulting from crossings at various speeds 14 (2023-04) 100162. doi:10.1016/j.dibe.2023.100162

    work page doi:10.1016/j.dibe.2023.100162 2023

  44. [44]

    Malekjafarian, F

    A. Malekjafarian, F. Golpayegani, C. Moloney, S. Clarke, A Machine Learning Approach to Bridge-Damage Detection Using Responses Mea- sured on a Passing Vehicle 19 (2019-01) 4035

    work page 2019

  45. [45]

    D. P. Kingma, J. Ba, Adam: A Method for Stochastic Optimization (2017). arXiv:1412.6980, doi:10.48550/arXiv.1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1412.6980 2017

  46. [46]

    Biewald, Experiment tracking with weights and biases, software avail- able from wandb.com (2020)

    L. Biewald, Experiment tracking with weights and biases, software avail- able from wandb.com (2020). URL https://www.wandb.com/

    work page 2020

  47. [47]

    G. Fan, J. Li, H. Hao, Vibration signal denoising for structural health monitoring by residual convolutional neural networks 157 (2020-06) 107651. doi:10.1016/j.measurement.2020.107651

    work page doi:10.1016/j.measurement.2020.107651 2020

  48. [48]

    K. He, X. Zhang, S. Ren, J. Sun, Deep Residual Learning for Image Recognition (2015-12-10). arXiv:1512.03385. 40

    work page internal anchor Pith review Pith/arXiv arXiv 2015

  49. [49]

    Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling

    J. Chung, C. Gulcehre, K. Cho, Y. Bengio, Empirical Evaluation of Gated Recurrent Neural Networks on Sequence Modeling (2014-12-11). arXiv:1412.3555

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  50. [50]

    J. Shi, H. Shi, J. Li, Z. Yu, Train-induced vibration response recon- struction for bridge damage detection with a deep learning methodology, Structures 64 (2024) 106496

    work page 2024

  51. [51]

    Attention Is All You Need

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, CoRR abs/1706.03762 (2017). arXiv:1706.03762. URL http://arxiv.org/abs/1706.03762 41

    work page internal anchor Pith review Pith/arXiv arXiv 2017