CoChat
Sign Up
Cited by 22Structural Health Monitoring TechniquesInfrastructure Maintenance and MonitoringConcrete Corrosion and Durability

Vibration-based structural health monitoring exploiting a combination of convolutional neural networks and autoencoders for temperature effects neutralization

Open access
Parziale, Marc, Lomazzi, Luca, Giglio, Marco, Cadini, Francesco·2022·Structural Control and Health Monitoring
DOI 10.1002/stc.3076
Open this paper in CoChat — free accountOpen PDF

Abstract

Damage diagnosis in the structural field (mechanical, civil, aerospace, etc.) is a topic of active development and research. In recent years, considerable enhancements in this field have been achieved mainly due to advances in sensor technologies, the evolution of signal processing algorithms, and the increase of computational power. As one of the main consequences, the amount of data recorded from the sensorial equipment has steadily grown in quantity and complexity. In addition to that, these data are almost always significantly affected by many factors, which are not only related to the presence of damages but, for instance, also to the environmental and operative conditions under which the structural system is working. In order to handle these challenges, in the last few years, new deep learning models have been proposed, based on deep and heterogeneous architectures, able to deal with big data, also containing intricate diagnostic features that are difficult to be extracted. With this aim, this paper proposes a new vibration-based structural diagnosis tool that exploits the power of convolutional neural networks (CNNs) to extract subtle damage-related features from complex transmissibility function (TF) spectra even in presence of potentially confounding temperature variations. The diagnostic algorithm stems from the coupling of a CNN with an unsupervised anomaly detection algorithm based on autoencoders (AEs) to neutralize the effects of temperature variations and increase the damage diagnosis accuracy. The proposed approach is demonstrated with reference to a simple, but realistic, numerical case study of a structural beam subjected to temperature changes.

AI snapshot

Written from the paper’s full text.

TL;DR

The paper couples a CNN that reads transmissibility-function spectrograms with per-element autoencoders to suppress temperature-induced false alarms in vibration-based structural health monitoring. In a numerical aluminium beam case study, the combined CNN-AE approach keeps false-alarm rates low across −30 °C to 100 °C, whereas the CNN alone degrades sharply at higher temperatures.

Key findings

  • A CNN trained only at 20 °C can localize and size stiffness-reduction damage, but temperature changes produce spurious damage predictions in healthy elements. For example, with 9% damage in element 8, the CNN predicted 2.5% at 70 °C versus 9% actual, and healthy element 19 exceeded 10%.
  • The CNN’s false-alarm rate rose above 40 °C and reached 109% at 100 °C, while the CNN-AE algorithm stayed mostly below 10% on average over the same range. At 100 °C, false alarms were 109% for CNN versus 4% for CNN-AE; at 90 °C, 81.6% versus 2.8%; at 80 °C, 46.3% versus 13.3%.
  • Missed-alarm rates were similar for both methods, with slight benefits for CNN-AE at temperature extremes. At 100 °C, missed alarms were 16.1% for CNN and 13.8% for CNN-AE; at −30 °C, 15.1% versus 10.7%.
  • The autoencoders reconstructed the undamaged CNN prediction–temperature relationship with medians near zero and negligible dispersion. The CNN also generalized to unseen damage extents of 15% and 30%, with prediction medians close to the ideal regression line.

Methods

  • A finite-element Timoshenko beam model was used: 1 m long, 25 mm × 10 mm aluminium cross-section, discretized into 20 elements. Damage was simulated as a stiffness reduction percentage, and temperature affected the Young’s modulus via E(T) = 8 × 10¹⁰ − 4 × 10⁷ × T. Ten transmissibility functions were computed from eleven FRFs over 0–2000 Hz, with artificial noise added.
  • The CNN, based on AlexNet, was trained on 15,000 spectrogram images resized to 227 × 227 pixels at a constant 20 °C, with state vectors giving damage location and extent. Validation used 3,000 new spectrograms. Training used stochastic gradient descent with momentum.
  • Twenty shallow overcomplete autoencoders, one per beam element with ten hidden neurons, were trained on undamaged-beam cases across −30 °C to 100 °C to reconstruct the CNN’s erroneous damage-extent versus temperature relationship. An anomaly index threshold of 0.85 and a CNN damage threshold of 8% were set by trial and error. Robustness was assessed with 1,000 damaged scenarios per temperature, sampling damage location, extent, and noise, plus a spatially linear temperature case from 45 °C to 75 °C.

Limitations

  • The paper states future work may address damage quantification under changing temperature fields and include other external factors such as external loads, as well as combining other algorithms or signal types for better damage characterization.
  • The study is a numerical case study using synthetic data from a finite-element model; no experimental validation is reported. This limitation is evident from the text.
  • Only temperature is treated as the confounding environmental factor, and the demonstration uses a single simple beam geometry and material. Damage is represented only as stiffness reduction, and the autoencoder approach assumes independent per-element training and operation. These limitations are evident from the described setup.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1002/stc.3076

Cites

47 of 47 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

7 of 16 cited sources checked · 3 supported · 0 flagged · accuracy not yet measured

  • Premisecitation

    Overcomplete AEs may learn a more generic representation of the input and provide better performances in various tasks than undercomplete AEs.

    “several studies suggested that overcomplete AEs, compared to undercomplete, may learn a more generic representation of the input and therefore provide better performances in various tasks.38,53”
    Quote in the paperSubstantial claim1 of 2 sources checked
    • cites [38]
    • cites [53] source

      The Introduction (p. 1) reports that overcomplete sparse coding outperformed bottleneck autoencoders on several measures of thumbnail-image compression. But the paper compares sparse coding with autoencoders, not overcomplete with undercomplete autoencoders; it does not establish that overcomplete autoencoders learn more generic representations or perform better across various tasks. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
  • Premisecitation

    Temperature is a potential confounding factor for SHM systems, especially large infrastructures, because temperature-induced variations in measured signals can have magnitudes similar to those generated by damages.

    “temperature is well known to be a potential confounding factor for SHM systems, especially when applied to large infrastructures, due to the variations induced on the measured signals, whose magnitudes, under certain circumstances, may be similar to those generated by damages.”
    Quote in the paperSubstantial claim1 of 3 sources checked
    • cites [19]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [20]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

    • cites [21] source

      The Experimental Setup and Results section reports temperature-induced changes in measured signal phase and amplitude, and the Concluding Remarks say these changes can cause false-positive damage detection in healthy structures. The copy does not establish that the signal variations are similar in magnitude to those caused by damage, or that the problem is especially pronounced in large infrastructure. (LLM, conf=0.84)

      Experimental — accuracy not yet measured
  • Premisecitation

    Transmissibility functions depend only on the location of the excitation load, not on its magnitude.

    “The main advantage of TFs is that they only depend on the location of the excitation load, but not on its magnitude.”
    Quote in the paperSubstantial claim1 of 1 sources checked
  • Premisecitation

    The choice of a 227×227 pixel input size is based on AlexNet having satisfactory performances with an input matrix of that size or very similar.

    “In particular, the choice of this specific size is based on the fact that the AlexNet 34 (and also similar CNN architectures) has been shown to offer satisfactory performances with an input matrix of that size or very similar, albeit for different types of problems with respect to the one presented in this work (e.g., image classification).”
    Quote in the paperNot a substantial claim2 of 2 sources checked
    • cites [34] source
    • cites [48] source

      Section 6 reports that AlexNet won the 2012 ImageNet image-classification competition, supporting the claim that it performed well on a different type of problem. It does not specify a 227×227-pixel input, a similar input size, or performance tied to that size (searched: “227” and AlexNet input dimensions). (LLM, conf=0.85)

      Experimental — accuracy not yet measured
  • Findingown data

    At T=70°C, for damage at element 8 with actual d%=9%, the CNN predicted d%=2.5%, far from the expected value.

    “the same does not happen for the 70 C scenario, where the predicted damage percentage is far from the expected one (d% CNN ¼ 2:5% with respect to an actual value of 9%)”
    Quote in the paperSubstantial claim
  • May combine two claims. 12
  • The quote only partly states this claim. 3
  • Classified as probably not a substantial claim. 4

  • May combine two claims.

    Premisecitation

    Transmissibility functions have a strong relationship with system damages and have been proposed as damage sensitive features.

    “TFs have been shown to have a strong relationship with system damages and, as such, they have been proposed as damage sensitive features in several studies in the literature.”
    Quote in the paperSubstantial claim2 of 3 sources checked
    • cites [18] source
    • cites [27] source

      Section 2.10 ('Other applications'; page unavailable) says several studies applied transmissibility to damage detection, supporting its use as a damage-sensitive feature. The discussion of Properties 1 and 2 shows that some mass or stiffness modifications leave the transmissibility matrix unchanged, so the paper does not support an unqualified strong relationship with system damage. (LLM, conf=0.83)

      Experimental — accuracy not yet measured
    • cites [28]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • May combine two claims.

    Findingown data

    The CNN generalizes to unseen damage extents (15% and 30% missing from training), with median predictions close to the ideal regression line.

    “The boxplots of Figure 9 show the satisfactory CNN performances even when fed with spectrograms corresponding to unseen damage scenarios (in terms of extent d%).”
    • Also stated in: § 3.1 | Convolutional neural network with transmissibility functions · plain
      “for both d% = 15% and d% = 30%, the medians of the CNN prediction distributions lie close to the ideal regression line (light blue).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    At T=20°C, for damage at element 8 with actual d%=9%, the CNN predicted d%=10.4%, and all undamaged elements remained below 5%.

    “for the 20 C scenario the damage prediction for the damaged element is still close enough to the real one (d% CNN ¼ 10:4% with respect to an actual value of 9%) and for all the undamaged elements remains below the 5%”
    Quote in the paperSubstantial claim

  • The quote only partly states this claim.

    Findingown data

    For undamaged beam elements, the median and dispersion of spurious CNN damage predictions increase as temperature increases across constant temperature fields from -30°C to 100°C.

    “The boxplots of Figure 12 show, for each temperature value considered, the distribution of the spurious fluctuations arising in the CNN prediction for the undamaged beam elements.”
    • Also stated in: § 3.1 | Convolutional neural network with transmissibility functions · hedged
      “The distributions median and dispersion values tend to increase as the temperature increases.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The use of unsupervised AEs guarantees better performance along the entire temperature range considered, allowing a more uniform trend for FA with lower percentage values.

    “The use of unsupervised AEs has demonstrated to be a powerful tool that guarantees better performance along the entire temperature range considered, allowing to have a much uniform trend for FA, with lower percentage values.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The CNN-AEs algorithm maintains a uniform false alarm rate across the temperature range, averaging below 10%.

    “On the other hand, the CNN-AEs algorithm (red circles) significantly mitigates the effects of temperature, keeping a uniform FA rate over the whole temperature range, remaining, on average, below 10%.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    For a spatially linear temperature profile from 45°C to 75°C, the CNN-AEs algorithm limits false alarm rate and maintains a missed alarm rate slightly lower than CNN alone.

    “As expected, the combined algorithm (CNN-AEs) is able to effectively limit the FA rate, maintaining a MA rate slightly lower than that obtained by employing the CNN only, showing the same behaviour seen for the constant temperature field.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). The quote only partly states this claim.

    Study design and conductnone given

    The dataset size of 15,000 spectrogram images was chosen by trial and error to obtain the desired algorithm accuracy at reasonable computational costs.

    “The choice of the dataset size is made by trial and error in order to obtain the desired algorithm accuracy at reasonable computational costs.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    In real applications or laboratory tests, transmissibility functions are obtained by collecting measurements from accelerometers and calculating the associated frequency response functions using appropriate estimators.

    “Note that, in real applications or laboratory tests the TFs are obtained by collecting measurements from accelerometers and by calculating the associated FRFs using appropriate estimators.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [47] source

  • May combine two claims.

    Interpretationreasoning

    The dataset for training CNNs with both damage and temperature variations would be very large and would never include all possible scenarios, limiting its effectiveness to a few specific cases with limited generalization.

    “The size of the dataset would be very large, given the large number of free parameters to be considered and, in any case, it would never be able to include all the possible scenarios, limiting its effectiveness just to a few, specific cases, with limited possibilities of generalization.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.

    Premisereasoning

    The selection of the proper deep neural network architecture is not always simple, and using an already existing one can avoid this process and potentially provide better results in terms of accuracy.

    “Indeed, the selection of the proper DNN architecture is not always a simple task and the use of an already existing one, characterized by many convolutional and pooling layers, can avoid this process and potentially provide better results in terms of accuracy.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Premisenone given

    The use of a CNN allows processing raw transmissibility functions and automatically extracting the most important features of the input.

    “Indeed, the use of a CNN allows processing raw TFs and automatically extracting the most important features of the input.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation

    It is not usually feasible to obtain datasets large enough for effective CNN construction from experiments; therefore, training data are synthetically generated with a numerical model.

    “Since it is not usually feasible to obtain datasets large enough for an effective construction of a CNN model from experiments, in this work, the common strategy of resorting to training data synthetically generated with a numerical model of the system is followed.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [31]

      Couldn’t be matched to a paper in Crossref, so it can’t be checked.

  • The quote only partly states this claim.

    Interpretationown data

    The spurious fluctuations in undamaged elements at T=70°C lead to a risk of false alarms.

    “leading to the risk of false alarms.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    When an element is actually damaged, the CNN outputs are not expected to be coherent with the d%_CNN-T_S relationship for the undamaged case, so differences will arise between AE inputs and outputs.

    “By doing so, when an element is actually damaged, the CNN outputs are not expected to be coherent anymore with the d% CNN n À T S n relationship for the undamaged case and so, some differences will arise between the AE inputs (d% CNN n À T S n ) and outputs (d% AE n , T AE n )”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Study design and conductown data

    Based on first results obtained in this work, both L2-based and sparsity-based regularization strategies are employed to avoid overfitting.

    “Based on the first results obtained in this work, not reported here for the sake of brevity, both the L2-based and sparsity-based regularization strategies are employed in this work to avoid overfitting.”
    Quote in the paperNot a substantial claim
16 of 16

  • The authors state this as an assumption.

    The temperature distribution is homogeneous within an individual element.

    “assuming homogeneous distribution of the temperature within an individual element”

  • The paper relies on this without support.

    The Young's modulus of the beam material depends on temperature.

    “given that the stiffness is strictly related to the Young's modulus, which, in turn, depends on the temperature”
2 of 2

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • A coupling method of CNN and autoencoder neutralized the influence of temperature change and improved damage detection accuracy on a limited dataset.

    “Parziale et al. [183] proposed a coupling method of a CNN and autoencoder to neutralize the influence of temperature change and improve the accuracy of damage detection on a limited dataset.”

    From Deep Learning for Structural Health Monitoring: Data, Algorithms, Applications, Challenges, and Trends (2023) · cites this paper as [183]

  • The problem of data temporal and spatial feature extraction can be solved using a combination of CNN and RNN methods.

    “The problem of data temporal and spatial feature extraction can be solved using a combination of the CNN and RNN methods, such as CNN with GRU [64,109,123], CNN with LSTM [27,178,179,180,181,182], CNN with Auto-encoder and [183], CNN with echo state networks [113], etc.”

    From Deep Learning for Structural Health Monitoring: Data, Algorithms, Applications, Challenges, and Trends (2023) · cites this paper as [183]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Claim of better performance across the entire temperature range conflicts with Table 6

    Major

    At 20 °C, Table 6 reports higher false-alarm and missed-alarm rates for CNN-AEs than for CNN alone: 7.3% versus 3% and 7.6% versus 6.2%, respectively. The combined algorithm therefore does not perform better across the entire range.

1 of 1

Cites

Flagged first, then by citations

1 of the 48 references couldn’t be matched to a known paper.

Research works printed but not linked: 5

  1. AvciO AbdeljaberO KiranyazS InmanD.Structural damage detection in real time: implementation of 1D convolutional neural networks for SHM applications.Conference Proceedings of the Society for Experimental Mechanics Series 2017. doi:10.1007/978‐3‐319‐54109‐9_6 · DOI
  2. SampaioRPC MaiaNMM RibeìroAMR SilvaJMM.Transmissibility techniques for damage detection.Proceedings of the International Modal Analysis Conference ‐ IMAC vol.2 2001.
  3. KullaaJ.Benchmark data for structural health monitoring. 7th European Workshop on Structural Health Monitoring EWSHM 2014‐2nd European Conference of the Prognostics and Health Management (PHM) Society 2014.
  4. RifaiS VincentP MullerX GlorotX BengioY.Contractive auto‐encoders: Explicit invariance during feature extraction. Proceedings of the 28th International Conference on Machine Learning ICML 2011 2011.
  5. 10.1007/978?3?642?35289?8_3 · DOI

Cited by

Sorted by citations

Related

Authors
Parziale, M., Lomazzi, L., Giglio, M., & Cadini, F.
Affiliations
Politecnico di Milano (IT)
Corresponding author
Cadini, Francesco
ORCID
Venue
Structural Control and Health Monitoring
Publisher
Wiley
Published
2022-08-23
Volume
29
Issue
11
ISSN
  • 1545-2255 (ISSN-L)
  • 1545-2263
DOI
10.1002/stc.3076
OpenAlex
W4292878042
Cited by
22
References
48
Citations by year
2026: 2 · 2025: 8 · 2024: 6 · 2023: 4
FWCI
2.08
Type
journal article
Language
en
Open access
Licence
Locations
  • Structural Control and Health Monitoring (publishedVersion)
  • Virtual Community of Pathological Anatomy (University of Castilla La Mancha) (submittedVersion)
Topics
  • Physical Sciences › Engineering › Civil and Structural Engineering › Structural Health Monitoring Techniques
  • Physical Sciences › Engineering › Civil and Structural Engineering › Infrastructure Maintenance and Monitoring
  • Physical Sciences › Engineering › Civil and Structural Engineering › Concrete Corrosion and Durability
Keywords
vibration-based structural health monitoring, damage diagnosis, convolutional neural networks, autoencoder, unsupervised anomaly detection, temperature variation, structural beams, deep learning

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Cited sources are checked by gpt-6-sol.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review