Physics-Informed Multimodal Bearing Fault Classification under Variable Operating Conditions using Transfer Learning

MD Manjurul Ahsan; Shivakumar Raman; Tasfiq E. Alam

REVIEW 4 major objections 4 minor 24 references

Physics-Informed Multimodal Bearing Fault Classification under Variable Operating Conditions using Transfer Learning

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A CNN that adds a physics-based penalty for physically implausible fault frequencies beats data-only training across operating conditions and reaches 98 percent accuracy on a second bearing dataset.

desk verdict A plausible but incremental bearing-fault classification recipe; the physics-loss sensitivity to variable speed is a real concern, and the supplied text is unreadable so the full paper needs refereeing. read the letter →

arxiv 2508.07536 v1 pith:POQGSAVJ submitted 2025-08-11 cs.LG

Tasfiq E. Alam , Md Manjurul Ahsan , Shivakumar Raman This is my paper · ORCID

classification cs.LG

keywords bearingfaultclassificationphysics-informedlossBPFOBPFItransferlearningmultimodalfusionrotatingmachineryvariableoperatingconditions

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

The paper sets out to show that a fault classifier for rolling-element bearings can be made more accurate and more portable by giving it a small amount of physics, not just data. The physics enters as a penalty term in the loss that rejects predictions inconsistent with the expected ball-pass frequencies of the outer and inner bearing races, computed from the bearing's geometry and shaft speed. The architecture is a convolutional network with a late fusion of vibration and motor-current signals, plus a physics-informed feature branch. On a benchmark bearing dataset the physics-informed model beats the same model without the physics term across multiple splits, and the layer-wise transfer strategy generalizes best to new operating conditions and machines, reaching roughly 98% accuracy on a held-out dataset. If the authors are right, this is a low-cost way to make deep fault diagnosis work under changing loads and speeds.

What carries the argument

The load-bearing object is the physics-informed loss, built from the BPFO and BPFI formulas. BPFO and BPFI are the periodic rates at which a defect on the outer or inner race encounters the rolling elements; they depend on the number of balls, ball diameter, pitch diameter, contact angle, and shaft speed. The loss asks the network's prediction to show energy near one of these expected fault frequencies and penalizes it when it does not. Around this loss, the model has a vibration branch, a motor-current branch, a late fusion step, and transfer-learning variants (TSFT, LAS, HFR), with LAS identified as the best transfer mode.

What would settle it

Take a bearing signal recorded during a smooth shaft-speed ramp. Compute BPFO/BPFI from the nominal speed and also from the instantaneous measured speed. If the physics-informed model trained with nominal-speed penalties scores worse than the identical model trained with instantaneous-speed penalties, or worse than no physics penalty at all, the central claim that the physics term helps under variable conditions fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that shaping the loss with characteristic bearing fault frequencies improves both accuracy and robustness compared with an otherwise identical CNN. The physics-informed loss uses BPFO (ball pass frequency, outer race) and BPFI (ball pass frequency, inner race), derived from bearing geometry and shaft speed, to penalize predictions that have no plausible fault-frequency content. The authors further report that layer-wise transfer learning (LAS) gives the best generalization across data splits, and that combining LAS with the physics-informed loss produces additional gains. Cross-dataset validation reaches up to 98% accuracy, and a statistical test supports the imp

Load-bearing premise

The entire physics penalty rests on the bearing geometry and shaft speed being known accurately at every inference; if the real instantaneous speed differs from the value used to compute BPFO/BPFI, the loss treats correct faults as impossible.

Editorial extensions

If this is right

A physics-informed loss can be added to an existing CNN classifier to improve fault accuracy without extra hardware.
Layer-wise adaptation is the recommended transfer strategy for moving models to new operating conditions.
The physical penalty and transfer learning combine for greater robustness than either one alone.
The same frequency-based penalty should extend to other rotating components with known characteristic defect frequencies.
The reported $p<0.01$ result suggests the improvement is consistent across splits, not a single lucky run.

Reading between the lines

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

The method's dependence on shaft speed implies an online speed-estimation branch is a natural next step if operating speeds are not fixed; otherwise the penalty may misfire during load transients.
A testable extension is to vary the bearing geometry parameters used in the loss and measure sensitivity; this would reveal how precisely the geometry metadata must be known.
If the penalty is kept at inference time, it could double as a confidence signal, flagging samples whose fault frequencies are ambiguous rather than forcing a nominal class.

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

Read this for what it is: a reasonable engineering paper in a mature field, not a breakthrough. The abstract describes a late-fusion multimodal CNN (vibration + current) with a physics penalty on BPFO/BPFI and three transfer-learning strategies, evaluated on Paderborn and KAIST. That combination is the new package, and it's a useful recipe for practitioners who want to move a bearing-fault classifier across machines without re-labeling everything. Credit where due: two public datasets, a statistical test with p<0.01, and a clear comparison of fine-tuning strategies. Those are concrete and reproducible if the code/data are released.

I can't size the errors, though. The supplied full text is corrupted to the point of unreadability—it even opens with a line from a different arXiv paper. That's almost certainly an extraction artifact, not a science problem, but it means I can only judge the abstract. On the abstract alone, the biggest conceptual risk is exactly the one you flagged: the physics loss is a function of bearing geometry and shaft speed, while the paper's motivation is variable operating conditions and cross-dataset transfer. If they use nominal or averaged speed instead of instantaneous speed, the penalty will mark correct predictions as implausible at off-nominal operating points, and the claimed gain over the data-only baseline could shrink or reverse. The KAIST transfer makes this worse unless the geometry is re-parameterized per target bearing. This is a real sensitivity, not a nitpick—but the full paper might address it, and I can't tell.

There are also the abstract-level gaps: no dataset sizes, class balance, error bars, or baseline architecture, so 'up to 98% accuracy' is not interpretable. The novelty is in the combination, not the components. This is for PHM/predictive-maintenance readers who want a cross-machine transfer recipe. It deserves a serious referee, because the claims are falsifiable and the topic is relevant. I'd send it with specific requests: show how speed is estimated per sample, give a sensitivity analysis on the physics-loss weight, and report per-class results on KAIST. Without clear handling of speed/geometry calibration, the central claim won't stand up. I wouldn't cite it in my own work until the full paper is out and the variable-speed handling is clear.

Referee Report

4 major / 4 minor

Summary. The paper proposes a physics-informed multimodal CNN for bearing fault classification, combining vibration and motor current signals with a late fusion architecture and a physics-informed loss based on BPFO/BPFI characteristic fault frequencies. It reports consistent improvements over a non-physics-informed baseline on the Paderborn dataset, evaluates three transfer-learning strategies (TSFT, LAS, HFR), and validates cross-dataset performance on the KAIST bearing dataset with up to 98% accuracy and p<0.01. However, the supplied full text is severely corrupted (mojibake), so the equations, experimental tables, ablation designs, and statistical procedures cannot be audited; only the abstract and a few fragmented formulas are recoverable.

Significance. If the claims hold, the work would provide a practically useful recipe for cross-machine bearing fault diagnosis, showing that a physics-informed loss grounded in bearing kinematics can combine with multimodal late fusion and transfer learning to improve generalization under variable operating conditions. The external KAIST validation, the use of standard BPFO/BPFI formulas, and the explicit reporting of statistical testing are strengths. The significance is undermined, however, by the inability to verify the central comparison, the physics-loss sensitivity to operating-speed accuracy, and the absence of any readable experimental detail.

major comments (4)

[Full text] The supplied full text is unreadable mojibake. No equation, table, ablation, or statistical procedure can be audited. The central claim — that the physics-informed model consistently outperforms the non-physics-informed baseline — therefore cannot be checked. This is a load-bearing deficiency, independent of the paper's technical merits.
[Physics-informed loss / BPFO-BPFI formulas] The displayed BPFO/BPFI formulas depend on shaft frequency f_s and bearing geometry. The abstract emphasizes 'variable operating conditions,' but nothing readable indicates whether f_s is the true instantaneous speed or a nominal/global value. If the latter, the loss will penalize correct predictions at off-nominal operating points, potentially reversing the claimed improvement. For the KAIST cross-dataset experiment, the target bearing geometry and speed must also be re-parameterized; the text gives no evidence that this was done.
[Statistical testing] The abstract reports p<0.01, but no test name, sample sizes, paired/unpaired design, or multiple-comparison correction is readable. With 'multiple data splits,' it is unclear what population the test refers to and whether the reported significance is between models or across splits. Without this information the p-value is not interpretable.
[KAIST cross-dataset validation] The abstract reports 'up to 98 percent accuracy' but gives no dataset size, class balance, availability of motor current signals, or description of how the late-fusion architecture and physics branch were adapted to KAIST. If KAIST labels were derived from spectral signatures closely aligned with BPFO/BPFI, the physics-informed loss could be rewarding label leakage rather than a robust physical prior. This concern must be addressed in a readable experimental section.

minor comments (4)

[Full text header] The body contains the arXiv header 'arXiv:2508.07544v1 [physics.chem-ph]', inconsistent with the stated paper ID and category. Please correct the compilation/watermark issue.
[Abstract] The abstract reports 'up to 98 percent accuracy' with no confidence intervals or error bars. Reporting the mean and standard deviation across splits would be more informative.
[Related work] The novelty of the 'novel physics-informed loss' should be positioned against prior bearing fault-diagnosis losses that use characteristic frequencies; the corrupted text makes this impossible to assess.
[Architecture description] Terms such as 'late fusion architecture' and 'dedicated physics-based feature extraction branch' need precise definitions and a figure/table specifying layer dimensions and fusion operations.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity identified; physics-informed loss is externally grounded and baseline comparisons are empirical.

full rationale

The paper's central physics-informed loss is based on BPFO/BPFI characteristic frequencies derived from bearing geometry and shaft speed, which are standard external formulas, not fitted from the target labels or from the baseline's performance. The claimed advantage over a non-physics-informed baseline is an empirical comparison across data splits and the KAIST cross-dataset benchmark, not a derivation that assumes the conclusion. The transfer-learning strategies are compared experimentally. The concern that variable operating conditions may invalidate the fixed-frequency penalty is a robustness/sensitivity issue, not circularity of the kind where a predicted quantity is identical to a fitted input by construction. The supplied full text is heavily corrupted (and includes a different arXiv identifier), so no equation-level reduction can be exhibited; the readable abstract contains no self-definitional, fitted-input, or self-citation-load-bearing step. Therefore no circular step is identified and the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

Only the abstract was readable. The central claim rests on the BPFO/BPFI physics formulas (external, standard), on dataset label reliability, and on the unreadable details of the loss function, architecture, and TL schedules. No fitted constants are visible in the abstract; an unknown penalty-weighting coefficient for the physics loss is expected, and the CNN carries the usual opaque hyperparameter budget. No invented entities appear.

free parameters (2)

Physics loss weighting coefficient(s)
The abstract describes a physics-informed loss that 'penalizes physically implausible predictions.' Such terms are weighted by hyperparameters, but no values or fitting procedure are given; the method section is unreadable.
CNN architecture and training hyperparameters
Kernel sizes, learning rates, batch sizes, fusion weights, and layer-freezing ratios for TSFT/LAS/HFR are not stated in the abstract and cannot be read in the corrupted full text.

assumptions (3)

domain assumption BPFO and BPFI characteristic fault-frequency formulas are the correct physical model for the fault classes in the datasets
The physics-informed loss is defined over these frequencies (abstract). The formulas are standard external knowledge, so they are not ad hoc, but their applicability to motor-current signatures is assumed.
domain assumption Paderborn and KAIST ground-truth labels and operating-condition metadata are reliable
Headline accuracies and the p<0.01 tests are computed against these labels; if labels encode the same spectral signatures the loss rewards, part of the physics gain would be circular.
domain assumption The transfer-learning strategies preserve the physics prior on the target domain
The claim that layer-wise adaptation plus physics modeling generalizes across datasets presupposes that BPFO/BPFI penalties remain valid at KAIST operating conditions.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Physics-Informed Multimodal Bearing Fault Classification under Variable Operating Conditions using Transfer Learning." pith.science (2026). https://pith.science/paper/POQGSAVJ

@misc{pith2026250807536,
  author       = {Pith},
  title        = {Pith review of: Physics-Informed Multimodal Bearing Fault Classification under Variable Operating Conditions using Transfer Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/POQGSAVJ}},
  note         = {Machine review of arXiv:2508.07536}
}

read the original abstract

Accurate and interpretable bearing fault classification is critical for ensuring the reliability of rotating machinery, particularly under variable operating conditions where domain shifts can significantly degrade model performance. This study proposes a physics-informed multimodal convolutional neural network (CNN) with a late fusion architecture, integrating vibration and motor current signals alongside a dedicated physics-based feature extraction branch. The model incorporates a novel physics-informed loss function that penalizes physically implausible predictions based on characteristic bearing fault frequencies - Ball Pass Frequency Outer (BPFO) and Ball Pass Frequency Inner (BPFI) - derived from bearing geometry and shaft speed. Comprehensive experiments on the Paderborn University dataset demonstrate that the proposed physics-informed approach consistently outperforms a non-physics-informed baseline, achieving higher accuracy, reduced false classifications, and improved robustness across multiple data splits. To address performance degradation under unseen operating conditions, three transfer learning (TL) strategies - Target-Specific Fine-Tuning (TSFT), Layer-Wise Adaptation Strategy (LAS), and Hybrid Feature Reuse (HFR) - are evaluated. Results show that LAS yields the best generalization, with additional performance gains when combined with physics-informed modeling. Validation on the KAIST bearing dataset confirms the framework's cross-dataset applicability, achieving up to 98 percent accuracy. Statistical hypothesis testing further verifies significant improvements (p < 0.01) in classification performance. The proposed framework demonstrates the potential of integrating domain knowledge with data-driven learning to achieve robust, interpretable, and generalizable fault diagnosis for real-world industrial applications.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

24 extracted references · 22 canonical work pages

[1]

Diagnosis of multiple faults in rotating machinery using ensemble learning

Udeme Ibanga Inyang, Ivan Petrunin, and Ian Jennions. Diagnosis of multiple faults in rotating machinery using ensemble learning. Sensors , 23(2):1005, 2023

work page 2023
[2]

A review of fault diagnosis methods for rotating machinery using infrared thermography

Rongcai Wang, Xianbiao Zhan, Huajun Bai, Enzhi Dong, Zhonghua Cheng, and Xisheng Jia. A review of fault diagnosis methods for rotating machinery using infrared thermography. Micromachines , 13(10):1644, 2022

work page 2022
[3]

Vibration analysis for machine monitoring and diagnosis: A systematic review

Mohamad Hazwan Mohd Ghazali and Wan Rahiman. Vibration analysis for machine monitoring and diagnosis: A systematic review. Shock and Vibration , 2021(1):9469318, 2021

work page 2021
[4]

Recent advances in the application of deep learning for fault diagnosis of rotating machinery using vibration signals

Bayu Adhi Tama, Malinda Vania, Seungchul Lee, and Sunghoon Lim. Recent advances in the application of deep learning for fault diagnosis of rotating machinery using vibration signals. Artificial Intelligence Review , 56(5):4667--4709, 2023

work page 2023
[5]

A survey on active deep learning: From model driven to data driven

Peng Liu, Lizhe Wang, Rajiv Ranjan, Guojin He, and Lei Zhao. A survey on active deep learning: From model driven to data driven. ACM Computing Surveys (CSUR) , 54(10s):1--34, 2022

work page 2022
[6]

When physics meets machine learning: A survey of physics-informed machine learning

Chuizheng Meng, Sam Griesemer, Defu Cao, Sungyong Seo, and Yan Liu. When physics meets machine learning: A survey of physics-informed machine learning. Machine Learning for Computational Science and Engineering , 1(1):20, 2025

2025
[7]

J. Zhao. A comprehensive review of deep learning-based fault diagnosis methods for rolling bearings. Advances in Physics: X , 2025

work page 2025
[8]

Habetler

Shen Zhang, Shibo Zhang, Bingnan Wang, and Thomas G. Habetler. Machine learning and deep learning algorithms for bearing fault diagnostics – a comprehensive review. arXiv preprint arXiv:1901.08247 , 2019

arXiv 1901

Show all 24 references

[9]

Shen et al

S. Shen et al. A physics-informed deep learning approach for bearing fault detection. Engineering Applications of Artificial Intelligence , 2021

2021
[10]

Hassannejad et al

R. Hassannejad et al. Adaptive wavelet-informed physics-based cnn for improved bearing fault diagnosis. International Journal of Prognostics and Health Management , 2025

2025
[11]

Bearing fault diagnosis under different working conditions based on domain adaptation using transferable features

Zhe Tong, Wei Li, Bo Zhang, and Meng Zhang. Bearing fault diagnosis under different working conditions based on domain adaptation using transferable features. arXiv preprint arXiv:1806.01512 , 2018

2018 arXiv
[12]

Asutkar et al

S. Asutkar et al. Deep transfer learning strategy for efficient domain-adaptive bearing fault diagnosis. Scientific Reports , 2023

2023
[13]

G. Xu, R. Guo, Z. Yin, and F. Zhang. Mcrcnet: A bearing fault diagnosis method for unknown faults based on transfer learning. Applied Sciences , 2025

2025
[14]

u chel, Christoph Schenck, Mohamed Selim, Karsten Starke, and Roland D. Br \

Christian Lessmeier, Jan Kurth, Klaus-Arno Schmidt, Bernd B \"u chel, Christoph Schenck, Mohamed Selim, Karsten Starke, and Roland D. Br \"a sel. Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benc...

2016
[15]

Harris and Michael N

Tedric A. Harris and Michael N. Kotzalas. Rolling Bearing Analysis: Advanced Concepts of Bearing Technology . CRC Press, Boca Raton, FL, 6 edition, 2021

2021
[16]

Randall and Jerome Antoni

Robert B. Randall and Jerome Antoni. Rolling element bearing diagnostics—a tutorial. Mechanical Systems and Signal Processing , 25(2):485--520, 2011

2011
[17]

Karniadakis, Ioannis G

George E. Karniadakis, Ioannis G. Kevrekidis, Lu Lu, Paris Perdikaris, Shandian Wang, and Liu Yang. Physics-informed machine learning. Nature Reviews Physics , 3:422--440, 2021

2021
[18]

Domain adaptation and transfer learning for intelligent fault diagnosis: A survey

Yu Yang et al. Domain adaptation and transfer learning for intelligent fault diagnosis: A survey. Mechanical Systems and Signal Processing , 190:110123, 2023

2023
[19]

A survey on transfer learning

Sinno J Pan and Qiang Yang. A survey on transfer learning. In IEEE Transactions on Knowledge and Data Engineering , 2009

2009
[20]

A bearing fault dataset for prognostics and diagnostics from kaist

Sang Lee et al. A bearing fault dataset for prognostics and diagnostics from kaist. IEEE Dataport , 2020

2020
[21]

Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark dataset for data-driven classification

Christian Lessmeier et al. Condition monitoring of bearing damage in electromechanical drive systems by using motor current signals of electric motors: A benchmark dataset for data-driven classification. In PHM Society European Conference , volume 3, pages 1--8, 2016

2016
[22]

Domain adaptation for fault diagnosis under variable working conditions using cnn and mmd

H Liu et al. Domain adaptation for fault diagnosis under variable working conditions using cnn and mmd. IEEE Transactions on Instrumentation and Measurement , 2021

2021
[23]

A transfer learning approach for cross-domain fault diagnosis of rolling bearings

Long Wen et al. A transfer learning approach for cross-domain fault diagnosis of rolling bearings. Mechanical Systems and Signal Processing , 135:106390, 2020

2020
[24]

Alam, Theodore Trafalis, and Pedro Huebner

Md Manjurul Ahsan, Tasfiq E. Alam, Theodore Trafalis, and Pedro Huebner. Deep mlp-cnn model using mixed-data to distinguish between covid-19 and non-covid-19 patients. Symmetry , 12(9):1526, 2020

2020

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Diagnosis of multiple faults in rotating machinery using ensemble learning

Udeme Ibanga Inyang, Ivan Petrunin, and Ian Jennions. Diagnosis of multiple faults in rotating machinery using ensemble learning. Sensors , 23(2):1005, 2023

work page 2023

[2] [2]

A review of fault diagnosis methods for rotating machinery using infrared thermography

Rongcai Wang, Xianbiao Zhan, Huajun Bai, Enzhi Dong, Zhonghua Cheng, and Xisheng Jia. A review of fault diagnosis methods for rotating machinery using infrared thermography. Micromachines , 13(10):1644, 2022

work page 2022

[3] [3]

Vibration analysis for machine monitoring and diagnosis: A systematic review

Mohamad Hazwan Mohd Ghazali and Wan Rahiman. Vibration analysis for machine monitoring and diagnosis: A systematic review. Shock and Vibration , 2021(1):9469318, 2021

work page 2021

[4] [4]

Recent advances in the application of deep learning for fault diagnosis of rotating machinery using vibration signals

Bayu Adhi Tama, Malinda Vania, Seungchul Lee, and Sunghoon Lim. Recent advances in the application of deep learning for fault diagnosis of rotating machinery using vibration signals. Artificial Intelligence Review , 56(5):4667--4709, 2023

work page 2023

[5] [5]

A survey on active deep learning: From model driven to data driven

Peng Liu, Lizhe Wang, Rajiv Ranjan, Guojin He, and Lei Zhao. A survey on active deep learning: From model driven to data driven. ACM Computing Surveys (CSUR) , 54(10s):1--34, 2022

work page 2022

[6] [6]

When physics meets machine learning: A survey of physics-informed machine learning

Chuizheng Meng, Sam Griesemer, Defu Cao, Sungyong Seo, and Yan Liu. When physics meets machine learning: A survey of physics-informed machine learning. Machine Learning for Computational Science and Engineering , 1(1):20, 2025

2025

[7] [7]

J. Zhao. A comprehensive review of deep learning-based fault diagnosis methods for rolling bearings. Advances in Physics: X , 2025

work page 2025

[8] [8]

Habetler

Shen Zhang, Shibo Zhang, Bingnan Wang, and Thomas G. Habetler. Machine learning and deep learning algorithms for bearing fault diagnostics – a comprehensive review. arXiv preprint arXiv:1901.08247 , 2019

arXiv 1901