Clustering-Based Classification of Polygonal Wheels in a Railway Freight Vehicle Using a Wayside System

Round 1

Reviewer 1 Report

The work presents a methodology for automatic classification of polygonal wheels in rail freight vehicles using accelerometers and signal processing techniques. Features are extracted using Continuous Wavelet Transform (CWT) and Principal Component Analysis (PCA) on Fast Fourier Transform (FFT) responses from sensors on the rail. The extracted features are then fused using Mahalanobis Distance to enhance sensitivity to abnormal cases. I have the following comments and questions regarding this paper:

1. Consider validating the methodology through experimental tests based on on-site measurements to further enhance its robustness and applicability in real-world scenarios.

2. Expand the testing scenarios to include multiple damages, different operating speeds, and various vehicle types to ensure the methodology's effectiveness across diverse conditions.

3. Currently, physics-informed neural networks have been widely applied in the field of deep learning, as evidenced by recent publications such as "Engineering Structures, 2023, 297: 117027" and "Computational Mechanics, 2021, 67: 207-230." These latest advancements should also be adequately highlighted in the introduction.

4. Explore the possibility of localizing polygonal wheels in multiple wagons of a freight train to broaden the scope of application and utility of the proposed methodology.

Minor editing of English language required.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

Thanks for inviting me to review the paper entitled “Clustering-based classification of polygonal wheels in a railway freight vehicle using a wayside system”. This paper proposes a machine learning-based method to classify different types of polygonal damage in freight vehicle wheels. Detecting railway wheel polygonal (OOR) defects is of interest, and this paper is structured well, but the following concerns must be addressed before accepting this paper.

(1)   The introduction is not comprehensive. If the authors pay attention to machinery fault diagnosis-related journals, e.g., Measurement and Mechanical System and Signal Processing, it can be easily found that there are many studies on railway wheel fault diagnosis, including OOR diagnosis. The related studies should be included to enrich the introduction.

(2)   In Ref. [https://doi.org/10.1016/j.measurement.2022.111268], many machine learning methods have been used to detect railway wheel OOR defects. The authors need to clarify their innovative point compared to the existing methods.

(3)   In this paper, simulated signals are used to demonstrate the feasibility of the proposed method, but Ref. [https://doi.org/10.1016/j.measurement.2023.112862] clearly demonstrates when using simulated signals to identify wheel OOR defects, many methods can achieve close to 100% accuracy because the simulation conditions are ideal.

(4)   Currently, many on-board methods have been published and applied to fault diagnosis of railway wheels. The author should state the advantages of the wayside monitoring methods.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The work is dedicated to the development of a methodology for detecting damaged wheels using accelerometers that measure vertical accelerations of the rail. Research on the condition of rolling stock wheels is relevant and practically valuable. There is currently a trend towards the widespread use of sensors for measuring accelerations not only of rolling stock elements but also of railway tracks. The accelerations of track elements objectively have a complex signal, the processing of which requires appropriate mathematical methods. The methods proposed by the authors will undoubtedly be of interest to researchers working on such issues.

To substantiate the hypothesis of detecting damaged wheels through acceleration recordings on the rail, the authors apply a comprehensive mathematical model. Certainly, comparisons with actual measurements need to be made, but the authors indicate in the conclusions that this will be the next stage of their research.

The reviewer has some recommendations and questions.

1. In the introduction, the authors write: "This proposed unsupervised methodology presents several novel aspects related to existing studies…" [16, 18, 28, 29, 35, 36]. Among them, only paper [16] is authored by other authors. It is desirable to compare the methodology not only with their own previous works but also with other studies.

2. In the text, the sensors are labeled A1-A4 (Line 154 and Fig. 2), and the wheel profiles are also labeled A1-A2 (Line 219). These labels are used simultaneously in Figs. 7 and 8, which may confuse the reader.

3. The authors claim that "with just two vertical accelerometers installed on each rail balanced results in terms of cost benefit were obtained.." However, the number "two accelerometers" is not substantiated in the manuscript, and it seems more logical to assume that there should be at least 3 sensors per rail.

4. The placement of the sensors shown in Fig. 2 requires additional explanations. Are the sensors specifically placed through the tie to be at a distance sleeper than the deflection length of the rail from the wheel, or is it better to install them at a greater distance?

5. The authors indicate that the railway track model is described in their previous works [17, 50] and provide a fragment of it in Fig. 3 for general demonstration. In paper [50], such an image of the model is absent, and paper [17] is not available in open access. I attempted to transfer the letter designations to the scheme, but I still have questions (see "Figure3.pdf" file).

6. The authors extensively examine the frequencies of oscillations caused by rolling stock, including wheel shape. Are the frequencies of inherent vibrations of railway track elements taken into account? Is the change in the spectrum of rail vibrations considered when there are sections with varying track stiffness? It is precisely the non-uniform track stiffness that is a significant factor influencing rail vibrations, not just the presence of geometric irregularities (Line 224-231).

7. The manuscript states: "Furthermore, an artificial noise equivalent to 5% of the amplitude is included in the numerical signal to provide a more realistic depiction of the measured rail response." The justification for the 5% value is not provided. Considering the previous question, this value may be significantly higher. It would be interesting to see the system's reaction in such a scenario.

8. The conclusions state: "This work proposes an automatic methodology based on machine learning…" However, "machine learning" is only mentioned in the introduction in the review of other works.

9. The conclusions state: "From an economic point of view, with only two accelerometers, the methodology can effectively classify in terms of dominant harmonic orders and defect amplitudes." However, this statement is not substantiated (see question 3).

10. The reference list includes 16 previous works involving the authors of this manuscript [17, 18, 27-29, 35-37, 40, 41, 43-46, 49, 50]. Indeed, this work is based on their previous research, so it doesn't make sense to repeat them. However, the number of self-citations can be reduced without losing the quality of material presentation. For example, the wheel motion model is described in works [44-46], paper [45] references paper [44], which can be excluded, and so on.

Comments for author File: Comments.PDF

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report

Dear Authors,

Firstly, allow me to commend you on your hard work and dedication towards enhancing the safety and reliability of railway transport through your innovative methodology for classifying polygonal wheels. The approach, leveraging a wayside system, showcases significant potential in addressing one of the prevalent issues in train wheel maintenance. While reviewing your manuscript, I have identified some areas that require minor revisions to ensure clarity, accuracy, and comprehensiveness. Please find my suggestions below:

1. Figure 4 lacks a clear definition of the Irregularity Level \(L_w\). It would be beneficial for readers if this definition could be incorporated into the manuscript.

2. The rendering of \(\psi_{\theta}\) in Equation (2) appears inconsistent with its representation on Line 210. Kindly make the necessary corrections to ensure uniformity.

3. In Table 1, the term "Noise Ratio" is used without a definition. Consider defining it or, preferably, using the more standard "Signal-to-Noise Ratio" in dB, which is commonly understood in the field.

4. The choice of the word "assessed" on Line 250 seems vague. A more precise term like "fixed" might be appropriate, along with a justification for its selection.

5. Equation (4) includes a notation (2) that may lead to confusion. Additionally, the use of \(\times\) for multiplication could potentially be misinterpreted as a cross-product. Please review and amend this notation for clarity.

6. There appears to be a discrepancy in Line 267 regarding whether it should be \(r\) or \(R_w\) in Equation (4) and its accompanying description. Please verify and correct as necessary.

7. The description from Lines 315-317 beginning with "In the first fusion step..." could benefit from improved clarity in its presentation.

8. Further elaboration on the application of the Mahalanobis distance within your methodology would enhance understanding. Including an equation or a detailed explanation would be valuable.

9. Establishing a baseline for comparison is crucial for contextualizing the efficacy of your proposed methodology. Consider incorporating a state-of-the-art work or a simplified version of your approach as a benchmark. Focusing on a key aspect of your pipeline, such as the use of CWT or sensor fusion, for this comparison could highlight the novelty and effectiveness of your work. An intra-cluster measure may serve as a suitable metric for this purpose.

I believe these revisions will strengthen the manuscript and better convey the significance of your contributions to the field. I look forward to reviewing the amended version of your work.

Best regards,

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The authors have addressed most of the concerns of the reviewer. The manuscript in its current form can be recommended for publication.

Reviewer 2 Report

Thanks for your kind response. This paper can be accepted.

Reviewer 3 Report

The authors provided thorough responses to all of my questions and implemented my recommendations. Some of the answers may be subject to debate, but this only demonstrates the relevance and scientific value of the article. The article has been revised, so I recommend it for publication in its current form.