A Review of Intelligent Device Fault Diagnosis Technologies Based on Machine Vision

Binrong Zhu; Guiran Liu

REVIEW 4 major objections 5 minor 25 references

A Review of Intelligent Device Fault Diagnosis Technologies Based on Machine Vision

T0 review · 4 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Transformer-based models achieve above-99% accuracy in mechanical fault diagnosis, a review contends.

desk verdict A review with a decent Transformer primer but citation integrity is poor enough that the headline accuracies are untraceable. read the letter →

arxiv 2412.08148 v1 pith:XTLPQU3I submitted 2024-12-11 cs.CV cs.AI

Guiran Liu , Binrong Zhu This is my paper · ORCID

classification cs.CVcs.AI

keywords Transformernetworksfaultdiagnosisdeeplearningmachinevisionself-attentionvibrationsignalsrollingbearingsreview

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This review argues that Transformer-based models have become the most accurate family of methods for intelligent mechanical fault diagnosis, with reported accuracies above 99 percent on public bearing and gearbox benchmarks. The load-bearing mechanism is self-attention, which lets the model weigh every part of a vibration signal against every other part and capture long-range dependencies that CNNs and RNNs miss. The paper assembles a comparative table showing Transformer variants outperforming convolutional and recurrent baselines, and it identifies the remaining bottlenecks: heavy labeled-data requirements, computational cost, and the gap between laboratory datasets and realistic aviation-engine data. If the review's reading of the literature is right, attention-based models are the natural baseline for future fault-diagnosis work, and the field's next gains will come from data efficiency, lightweight design, and robustness rather than from further architectural novelty alone.

What carries the argument

The central object is the Transformer's scaled dot-product self-attention, along with positional encoding and the Vision Transformer's patch-embedding trick. Self-attention computes a weighted sum over all sequence positions, letting each vibration-sample position attend to every other position in parallel; positional encoding preserves order when recurrence is removed. The review also relies on a preprocessing pathway that converts one-dimensional vibration signals into two-dimensional time–frequency images, so that image-oriented Transformer variants such as ViT and Swin can be applied. This machinery is what the paper credits for the above-99-percent accuracies and for the claimed superiority over CNNs and RNNs.

What would settle it

Check references [21] and [22] against the text: if [21] actually describes a wavelet-transform hybrid rather than a time-series transformer, and [22] is a general review rather than a proposed TAR joint-extraction method, then the review's accuracy comparisons for those rows are unsupported. A direct re-run of the named models on the same datasets under a fixed protocol would also settle whether the above-99-percent Transformer advantage reproduces.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that the Transformer architecture, originally built for language, transfers to vibration-based fault diagnosis and outperforms established deep-learning baselines. The review reports that methods built around self-attention, including TST, Diagnosisformer, SiT, and TAR, reach accuracies of 98.63 to 99.90 percent on three public laboratory bearing and gearbox datasets, while conventional models such as BPNN, DBN, 1DCNN, LSTM, and ResNet18 plateau below them in the same table. On the aviation-engine benchmark, the same Transformer method drops to about 71 percent, which the paper interprets as evidence that real-world machine fault data remains a harder and more representative test. The paper's conclusion is that Transformer-based models excel at extracting and recognizing fault-relevant patterns from complex sensor data, but their practical deployment is limited by data hunger, computation cost, and the controlled conditions of most benchmarks.

Load-bearing premise

The review's central account rests on the assumption that every cited study exists and that the accuracies in Table 1 are real, correctly measured, and correctly attributed to the named Transformer methods.

Editorial extensions

If this is right

If the reported accuracies hold, Transformer variants should replace CNNs and RNNs as the default baselines for vibration-based fault classification on public datasets.
The above-99-percent ceiling on laboratory benchmarks implies that further gains in published accuracy are less informative than progress on data efficiency, noise robustness, and cross-condition generalization.
The much lower accuracy on the aviation-engine benchmark implies that real-world fault diagnosis will need transfer learning, few-shot methods, or synthetic data before attention models reach deployment quality.
The paper's proposed hybrid CNN–Transformer direction follows directly: CNNs capture local transients while self-attention captures global context, and both are needed for short-sequence signals.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

A cautious reader should treat the accuracy ranking in Table 1 as provisional: the underlying studies likely differ in train/test splits, signal length, noise conditions, and number of fault classes, so the numbers may not be directly comparable even if every citation is genuine.
If any of the cited sources turn out to be misattributed, the review's comparative claim would need to be re-established from primary sources; the qualitative argument that self-attention helps long-range vibration pattern recognition could still survive independently of any single accuracy figure.
A natural testable extension is to benchmark the named Transformer variants against a shared protocol on the same datasets, including noisy and variable-speed conditions, to see whether the above-99-percent ordering reproduces outside each paper's original setup.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Colleague,

Here's my read on this review of Transformer-based fault diagnosis. Punchline: it's a review with no original results, and the reference problems are bad enough that the reported accuracies cannot be trusted. The paper would need a complete rework of its literature summary to be salvageable.

What it does well: the early sections give a reasonable primer on Transformer components (self-attention, FFN, positional encoding) and on ViT/DeiT. The dataset descriptions (CWRU, XJTU-SY, UCONN, HIT) are useful for readers new to the area. The future-directions list—lightweight models, few-shot learning, CNN hybridization—is standard but not wrong.

The soft spots are not minor. Table 1 is the heart of the survey's empirical claim: Transformer models outperform CNNs/RNNs with >99% accuracy on public benchmarks. But the citations attached to those numbers do not line up. Reference [21], credited for the Time-Series Transformer (TST), is a wavelet-transform hybrid model paper. Reference [22], credited for the Transformer-ResNet fusion (TAR), is a review of power-transformer fault diagnosis. Reference [24] for Diagnosisformer is a generic predictive-maintenance paper, and [25] for SiT is a generic vibration-monitoring review. The text also disagrees with the table (TST accuracy 99.72% for 4 classes vs. 98.63% in the table). Several references have truncated author names and DOIs that point to unrelated journals. There is no methodology section explaining how references were selected or numbers verified. So the central comparative claim is unsupported by the evidence provided.

That said, the paper does not appear to be a deliberate fraud in how it reasons—the narrative is coherent and the Transformer exposition is mostly accurate. But the citation failures are load-bearing. A survey that cannot trace its own key results is not a trustworthy entry point.

My recommendation: desk reject. The authors should redo the reference verification and ideally conduct a systematic review with a reproducible search protocol. A reader looking for a quick overview of Transformer basics might find the first three sections mildly useful, but I wouldn't cite or build on it.

Hope this helps.

Referee Report

4 major / 5 minor

Summary. This manuscript is a narrative review of Transformer-based methods for intelligent fault diagnosis of mechanical equipment. It describes the Transformer architecture, its self-attention mechanism and position encoding, reviews vision Transformer variants such as ViT and DeiT, and then surveys applications to bearing and gearbox fault diagnosis. The paper presents a table of classification accuracies for several methods across four public datasets (CWRU, XJTU-SY, UCONN, HIT) and concludes with challenges and future research directions, including few-shot learning, lightweight models, and CNN–Transformer hybrids.

Significance. A well-executed review of this rapidly growing area would be valuable, and the paper correctly identifies important challenges such as data scarcity, computational cost, and limited deployment to resource-constrained settings. However, the review lacks the basic apparatus of a scholarly survey: there is no search protocol, inclusion criterion, or evidence synthesis method. More seriously, the central empirical claim—that Transformer models achieve accuracies above 99% on the named benchmarks—is not traceable to the cited literature because the cited references do not correspond to the named methods. The paper provides no reproducible artifacts, so its value depends entirely on the accuracy of its literature account, which is not currently supported.

major comments (4)

[III.C, Table 1] The text states that TST achieved fault diagnosis accuracies of 98.63% (10 classes), 99.72% (4 classes), 99.78%, and 99.51%, respectively, but Table 1 lists 98.63% for CWRU, 99.78% for XJTU-SY, and 99.51% for UCONN, with no 99.72% entry and no class-count information. This internal inconsistency makes the headline accuracy claims ambiguous and prevents the reader from verifying which reported number corresponds to which experimental condition.
[III.C and References [21], [22], [24], [25]] The references cited for the key methods do not match the described methods. Reference [21], cited for Jin et al.'s Time-Series Transformer (TST), is titled "A hybrid model for bearing fault diagnosis with wavelet transform and machine learning," which describes a wavelet-based hybrid model, not a time-series transformer. Reference [22], cited for Hou et al.'s TAR, is a review of power-transformer fault diagnosis. Reference [24], cited for Diagnosisformer, is a general predictive-maintenance paper. Reference [25], cited for SiT, is a vibration-monitoring review. Consequently, the accuracy entries for TST, TAR, Diagnosisformer, and SiT in Table 1 have no verifiable source in this reference list.
[I and II] The paper claims to provide a comprehensive review but gives no literature search strategy, inclusion/exclusion criteria, or quality assessment. This omission prevents the reader from reconstructing how the surveyed papers and their reported accuracy numbers were selected, checked, or synthesized. For a review whose central claim is an empirical statement about the literature, the absence of a methodology is a load-bearing weakness.
[III.C, Table 1] The HIT dataset results, including the TST accuracy of 71.07% in Table 1, are presented as establishing "a new benchmark for validating mechanical fault diagnosis methods," but no citation or description of the HIT dataset source is provided. The only mention of HIT is in Figure 6's caption, which lists it alongside the other datasets without a reference. The benchmark claim is therefore unsubstantiated.

minor comments (5)

[III.A, Equations] The equation numbering is inconsistent: Equation (1) is used for both the FFN computation and the position-encoding formulas, and the position-encoding expressions are each labeled (1). The equations should be numbered sequentially and their mathematical notation typeset cleanly.
[III.A, Figures] Figure 2 is used twice with different captions (once for the RNN architecture and once for the original Transformer structure), and Figure 6 is referenced both as "Data collection equipment" and in the later text with a duplicated number. Figure numbers and captions need to be unique and consistent throughout.
[References] Several references have truncated author names (e.g., "iu, Z." in [3], "ang, J." in [4], "Hen, L." in [10]) and inconsistent DOI or page formatting. These bibliographic errors should be corrected before any resubmission.
[Keywords] The keyword list uses a full-width semicolon and inconsistent capitalization ("Fault Diagnosis；Vibration Analysis；Deep Learning；"). The keywords should be separated with standard semicolons and normalized.
[I, Reference [4]] The statement that mechanical equipment failures contribute to approximately 38% of major accidents and economic losses in industrial production cites reference [4], a general machine-learning book, without a specific page or study. A proper primary source is needed for this statistic.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the survey makes no fitted predictions or self-cited uniqueness claims; citation mismatches are an evidence problem, not a circular one.

full rationale

This is a literature review, not a derivation. It presents no fitted parameters, no equations whose outputs are defined by their inputs, no benchmark predictions generated from the cited models, and no load-bearing self-citations; the two authors' own prior work does not appear among the references supporting the survey's claims. The central empirical assertion that Transformer models reach above 99% accuracy is an attribution to external papers (e.g., refs [21], [22], [24], [25]). Even if that attribution is wrong because the listed titles describe different methods, that is a verifiability and correctness defect, not circularity: the survey's claims are not equivalent to its inputs by construction. The reference-title mismatches should be raised in a reliability review, but they do not satisfy any circularity pattern. Therefore the circularity score is 0.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

As a review, the paper has no free parameters and no invented entities. The central assumption is the validity of the cited literature, which is not satisfied based on the citation mismatches and likely fabrication.

assumptions (1)

domain assumption The cited papers exist and report the accuracy values presented in Table 1.
Section III.C relies on references [20]-[25] for TST, TAR, Diagnosisformer, and SiT. Several of these references are mismatched with the described methods, and some references contain truncated author names and DOIs that lead to unrelated journals, indicating the references may be fabricated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of A Review of Intelligent Device Fault Diagnosis Technologies Based on Machine Vision." pith.science (2026). https://pith.science/paper/XTLPQU3I

@misc{pith2026241208148,
  author       = {Pith},
  title        = {Pith review of: A Review of Intelligent Device Fault Diagnosis Technologies Based on Machine Vision},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XTLPQU3I}},
  note         = {Machine review of arXiv:2412.08148}
}

read the original abstract

This paper provides a comprehensive review of mechanical equipment fault diagnosis methods, focusing on the advancements brought by Transformer-based models. It details the structure, working principles, and benefits of Transformers, particularly their self-attention mechanism and parallel computation capabilities, which have propelled their widespread application in natural language processing and computer vision. The discussion highlights key Transformer model variants, such as Vision Transformers (ViT) and their extensions, which leverage self-attention to improve accuracy and efficiency in visual tasks. Furthermore, the paper examines the application of Transformer-based approaches in intelligent fault diagnosis for mechanical systems, showcasing their superior ability to extract and recognize patterns from complex sensor data for precise fault identification. Despite these advancements, challenges remain, including the reliance on extensive labeled datasets, significant computational demands, and difficulties in deploying models on resource-limited devices. To address these limitations, the paper proposes future research directions, such as developing lightweight Transformer architectures, integrating multimodal data sources, and enhancing adaptability to diverse operational conditions. These efforts aim to further expand the application of Transformer-based methods in mechanical fault diagnosis, making them more robust, efficient, and suitable for real-world industrial environments.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

25 extracted references · 20 canonical work pages

[21]

Sun, W., & Liu, H. (2022). A hybrid model for bearing fault diagnosis with wavelet transform and machine learning. IEEE Transactions on Industrial Informatics, 19(3), 1847–1855. https://doi.org/10.1109/TII.2022.3107929

work page arXiv 2022
[22]

Chen, Z., & Wang, S. (2022). Fault diagnosis of transformers using vibration and acoustic signals: A review. Electric Power Components and Systems, 50(12), 1189–1201. https://doi.org/10.1080/15325008.2022.2047539

work page arXiv 2022
[24]

Li, X., & Zhou, P. (2024). Predictive maintenance and fault detection using deep learning techniques. Neural Computing and Applications, 36(7),2319–2331.https://doi.org/10.1007/s00542-023-07089-4

work page doi:10.1007/s00542-023-07089-4 2024
[25]

Zhang, Q., & Wang, R. (2023). The use of vibration monitoring for industrial fault diagnosis: Methods and applications. Journal of Vibration Engineering, 29(2), 156-167. https://doi.org/10.1134/S1050448017020179

work page doi:10.1134/s1050448017020179 2023
[1]

(2023) Convolutional Neural Network-Based Transformer Fault Diagnosis Using Vibration Signals

Liu, Z., Davari, P., & Yang, J. (2023) Convolutional Neural Network-Based Transformer Fault Diagnosis Using Vibration Signals. Sensors,23(10):4781–4795.https://doi.org/10.3390/s23104781

doi:10.3390/s23104781 2023
[2]

Sahu, S., & Satapathy, P. K. (2023) A novel time–frequency Transformer based on self–attention mechanism and its application in fault diagnosis of rolling bearings. Mechanical Systems and Signal Processing,177:109028.https://doi.org/10.1016/j.ymssp.2023.109028

work page arXiv 2023
[3]

(2023) Transformer-Based Deep Learning for Vibration Fault Diagnosis

iu, Z., Yang, J., & Davari, P. (2023) Transformer-Based Deep Learning for Vibration Fault Diagnosis. In: Proceedings of the International Conference on Mechanical Engineering. Beijing, China. pp. 98-104. https://doi.org/10.1109/ICME52516.2023.101462

work page arXiv 2023
[4]

(2023) Advanced Machine Learning Techniques for Mechanical Systems

ang, J., & Liu, Z. (2023) Advanced Machine Learning Techniques for Mechanical Systems. Academic Press, San Diego. https://doi.org/10.1016/B978-0-12-823310-3.00001-0

work page doi:10.1016/b978-0-12-823310-3.00001-0 2023

Show all 25 references

[5]

(2023) Fault Diagnosis Methods for Mechanical Systems

iu, Z. (2023) Fault Diagnosis Methods for Mechanical Systems. In: Yang, J., & Liu, Z. (Eds.), Advanced Machine Learning Techniques for Mechanical Systems. Academic Press, San Diego. pp. 125-150. https://doi.org/10.1016/B978-0-12-823310-3.00008-3

2023 doi
[6]

(2023)Multi-Scale Transformer Network for Fault Diagnosis of Gearbox

Zhang, T., &Xu,W. (2023)Multi-Scale Transformer Network for Fault Diagnosis of Gearbox. Journal of Sound and Vibration, 529: 116864. https://doi.org/10.1016/j.jsv.2023.116864

2023
[7]

(2022) Deep Learning Based Transformer for Rolling Element Bearing Fault Diagnosis

Gao, Y., Chen, H., & Li, Y. (2022) Deep Learning Based Transformer for Rolling Element Bearing Fault Diagnosis. Mechanical Systems and Signal Processing, 168: 108663. https://doi.org/10.1016/j.ymssp.2022.108663

2022
[8]

(2022) Transformer-Based Deep Learning Approaches for Mechanical Fault Detection

Wang, H., & Zhang, L. (2022) Transformer-Based Deep Learning Approaches for Mechanical Fault Detection. In: Proceedings of the International Conference on Intelligent Systems and Engineering. Xi'an, China.pp.112-117.https://doi.org/10.1109/ISE53527.2022.9876543

2022
[9]

Application of Transformer Models in Mechanical Fault Diagnosis

Khan, M., 2023. Application of Transformer Models in Mechanical Fault Diagnosis. ResearchGate. https://www.researchgate.net/publication/372925641

2023
[10]

Machine Learning and Transformers in Engineering Diagnostics

Hen, L., 2022. Machine Learning and Transformers in Engineering Diagnostics. IEEE Xplore. https://ieeexplore.ieee.org/document/9521456

2022
[11]

Huang, T., Wang, Z., Li, J., & Zhang, S. (2024). Research on a Transformer Vibration Fault Diagnosis Method Based on Time-Shift Multiscale Increment Entropy and CatBoost. Entropy, 26(9), 721. https://doi.org/10.3390/e26090721

2024 doi
[12]

Patel, S., & Yadav, R. (2023). Dynamic Fault Diagnosis for Rotating Machinery Using Time–Frequency Analysis. Journal of Vibration and Control, 29(10), 1983–1995. https://doi.org/10.1177/10775463221116391

2023 doi
[13]

Zhang, L., & Chen, Z. (2022). A novel hybrid method for power transformer fault diagnosis. Electric Power Systems Research, 199, 107374.https://doi.org/10.1016/j.epsr.2021.107374

2022
[14]

IEEE Access, 11, 3271-3280

Liu,W.,&Wang,J.(2023).Theapplicationofdeeplearningtechniques in industrial fault diagnosis. IEEE Access, 11, 3271-3280. https://doi.org/10.1109/ACCESS.2023.3045376

2023
[15]

Zhang, X., & Li, H. (2023). A review on the fault detection systems for industrial equipment. Industrial Electronics Journal, 37(6), 542–556. https://doi.org/10.1109/JIE.2023.3570298

2023
[16]

Wang, Y., Zhang, Y., & Huang, D. (2022). Transformer fault detection using vibration signals and machine learning. Sensors, 22(7), 2534. https://doi.org/10.3390/s22072534

2022 doi
[17]

Smith, R., & Clark, P. (2023). Data-driven fault diagnosis for wind turbines based on vibration monitoring. Renewable Energy, 205, 1302–1313.https://doi.org/10.1016/j.renene.2023.03.038

2023 doi
[18]

Lee, T., & Chen, Y. (2023). Intelligent fault diagnosis for mechanical systems using ensemble learning. Mechanical Systems and Signal Processing,177,108775.https://doi.org/10.1016/j.ymssp.2022.108775

2023
[19]

Yang, Z., & Li, B. (2022). Multiscale entropy based fault detection for gearboxes. Journal of Sound and Vibration, 524, 116833. https://doi.org/10.1016/j.jsv.2022.116833

2022
[20]

Zhao, Q., & Zhang, J. (2023). Vibration fault diagnosis in high-speed machinery: Techniques and applications. Journal of Mechanical Engineering Science, 237(6), 2224–2237. https://doi.org/10.1177/09544054231155631

2023 doi
[23]

IEEE Transactions on Vehicular Technology, 72(4), 3428–3437

Yang,M.,&Sun,J.(2023).Faultdiagnosisinautomotivesystemsusing time-domain analysis and machine learning. IEEE Transactions on Vehicular Technology, 72(4), 3428–3437. https://doi.org/10.1109/TVT.2023.3212019

2023

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.

[1] [21]

Sun, W., & Liu, H. (2022). A hybrid model for bearing fault diagnosis with wavelet transform and machine learning. IEEE Transactions on Industrial Informatics, 19(3), 1847–1855. https://doi.org/10.1109/TII.2022.3107929

work page arXiv 2022

[2] [22]

Chen, Z., & Wang, S. (2022). Fault diagnosis of transformers using vibration and acoustic signals: A review. Electric Power Components and Systems, 50(12), 1189–1201. https://doi.org/10.1080/15325008.2022.2047539

work page arXiv 2022

[3] [24]

Li, X., & Zhou, P. (2024). Predictive maintenance and fault detection using deep learning techniques. Neural Computing and Applications, 36(7),2319–2331.https://doi.org/10.1007/s00542-023-07089-4

work page doi:10.1007/s00542-023-07089-4 2024

[4] [25]

Zhang, Q., & Wang, R. (2023). The use of vibration monitoring for industrial fault diagnosis: Methods and applications. Journal of Vibration Engineering, 29(2), 156-167. https://doi.org/10.1134/S1050448017020179

work page doi:10.1134/s1050448017020179 2023

[5] [1]

(2023) Convolutional Neural Network-Based Transformer Fault Diagnosis Using Vibration Signals

Liu, Z., Davari, P., & Yang, J. (2023) Convolutional Neural Network-Based Transformer Fault Diagnosis Using Vibration Signals. Sensors,23(10):4781–4795.https://doi.org/10.3390/s23104781

doi:10.3390/s23104781 2023

[6] [2]

Sahu, S., & Satapathy, P. K. (2023) A novel time–frequency Transformer based on self–attention mechanism and its application in fault diagnosis of rolling bearings. Mechanical Systems and Signal Processing,177:109028.https://doi.org/10.1016/j.ymssp.2023.109028

work page arXiv 2023

[7] [3]

(2023) Transformer-Based Deep Learning for Vibration Fault Diagnosis

iu, Z., Yang, J., & Davari, P. (2023) Transformer-Based Deep Learning for Vibration Fault Diagnosis. In: Proceedings of the International Conference on Mechanical Engineering. Beijing, China. pp. 98-104. https://doi.org/10.1109/ICME52516.2023.101462

work page arXiv 2023

[8] [4]

(2023) Advanced Machine Learning Techniques for Mechanical Systems

ang, J., & Liu, Z. (2023) Advanced Machine Learning Techniques for Mechanical Systems. Academic Press, San Diego. https://doi.org/10.1016/B978-0-12-823310-3.00001-0

work page doi:10.1016/b978-0-12-823310-3.00001-0 2023