Determination of the Condition of Railway Rolling Stock Using Automatic Classifiers

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This work presents a study on the application of machine learning techniques to vibration signals acquired from a railway bogie test bench, a topic of significant importance for improving railway maintenance and safety. However, several issues require attention before the manuscript is finalized. Specifically:

1. The abstract should be rewritten as a concise single paragraph to better convey the study's objectives, methods, and key findings.
2. The format of Figure 1 needs to be standardized to maintain consistency with other figures in the manuscript.
3. The table format is incorrect and should be adjusted to adhere to academic standards, ensuring clarity and readability.
4. The author should elaborate further on the novelty and uniqueness of the research in the preface to highlight its contributions to the field.
5. The manuscript currently resembles a project introduction or product manual, lacking the rigorous academic style expected of a research paper. Language and content should be refined to align with academic norms, focusing on theoretical and methodological details rather than descriptive or explanatory text.

Author Response

This work presents a study on the application of machine learning techniques to vibration signals acquired from a railway bogie test bench, a topic of significant importance for improving railway maintenance and safety. However, several issues require attention before the manuscript is finalized. Specifically:

We would like to express our sincere appreciation for their valuable comments, which have led us to modify the article accordingly. Thus, the improved version is not only the result of the authors' work, but also of the reviewers' hard work. All changes made as a result of the reviewers' comments are shown in blue in the revised article.

1. The abstract should be rewritten as a concise single paragraph to better convey the study's objectives, methods, and key findings.

Thank you very much for your comment. According to your recommendation, the abstract has been rewritten into a single paragraph, which reads as follows:

Efficiency of maintenance is a key point in systems of railway transport, in order to avoid catastrophic accidents. Therefore, having a method that allows for the early detection of defects in critical elements, is a crucial for both increasing the availability of rolling stock and reducing maintenance costs. The main contribution of this work is the proposal of a methodology for analysing vibration signals. Obtained from a bogie axle in a test bench, vibrating signals are decomposed in intrinsic functions and subsequently apply each function classical signal processing techniques. Finally by means of decision trees to characterise the status of the axle, achieving excellent results.

2. The format of Figure 1 needs to be standardized to maintain consistency with other figures in the manuscript.

Thank you very much for your appreciation. Figure 1 has been maintained in a format that differs slightly from the other figures in order to illustrate the functionality of the EMD method. However, should it be deemed necessary, the figure could be altered in its entirety or in part.

3. The table format is incorrect and should be adjusted to adhere to academic standards, ensuring clarity and readability.

Thank you very much for your appreciation. According to your recommendation the tables have been revised in accordance with the recommendations, with a view to aligning the presented model.

4. The author should elaborate further on the novelty and uniqueness of the research in the preface to highlight its contributions to the field.

Thank you very much for your comment. Following your recommendation to clarify the objectives of the work as well as its novelty, two new sections have been incorporated into the introduction:

1.5. Approaches and challenges

About other approaches that have been taken to achieve the goal of optimising the analysis of incoming data, Gafni et al. [45] identify Federated Learning as a tool for the future to, among other things, use information received from different sources which, in many cases, may not be fully reliable and/or relevant, to a certain extent, Hu et al. [46] are associated with not only different data sources, but also different research aims and the improvement of data. The amount of data gives rise to a number of challenges, but it also creates an area of interest in the form of decentralised data, as has been pointed out by Yang et al. [47], but the quality of the information, and thus the effectiveness of the analysis, is affected by the use of data from different sources and the methods of data collection.

1.6. Aim of the study

The present work explores the use of decision trees combined with EMD decomposition as an alternative method for classifying and categorising the condition of railway rolling stock. This is achieved analysing the different stages of deterioration of the components, with the goal of determining the initial state of mechanical failure and improving general predictive maintenance, which offers results comparable to those obtained by Ruiz Torres et al. [48], whose work was based on the use of Support Vector Machine, and Bustos et al. [49], who uses EMD and K-Nearest Neighbour to perform classification under similar conditions.

The methodology applied in this work is frequently used in bearing condition analysis, as evidenced by the extensive literature referenced in this introduction. This literature mainly concerns the methodology used for rolling bearings and has been presented more broadly in terms of both elapsed time and range of methodologies.

 

5. The manuscript currently resembles a project introduction or product manual, lacking the rigorous academic style expected of a research paper. Language and content should be refined to align with academic norms, focusing on theoretical and methodological details rather than descriptive or explanatory text.

Thank you very much for your hint. Thanks to your observations, the work has undergone a series of modifications and the introduction of new information with the aim of aligning it with the prevailing academic standards. These standards have been highlighted in blue for the purpose of improved identification.

Reviewer 2 Report

Comments and Suggestions for Authors

1. please improve resolutions of all figure and tables.
2. please add ref. for all equations and add latest ref. post 2020.
3. please put accelerometer specification as a table , what method was used to mount it on test bed?
4. which type of chamber was used for test rig?? how you consider resonance??
5. what is significance of depth e in line number 247??
6. please add a section about role of vibration as condition monitoring tool?
7. please add structure of paper at end of introduction section, how was figure 5 decomposition achieved??

Author Response

We would like to express our sincere appreciation for their valuable comments, which have led us to modify the article accordingly. Thus, the improved version is not only the result of the authors' work, but also of the reviewers' hard work. All changes made as a result of the reviewers' comments are shown in blue in the revised article.

1. please improve resolutions of all figure and tables.

Thank you for your hint. According to your suggestion the resolution of tables and figures has been enhanced to facilitate optimal visualisation. The following example is provided for illustrative purposes:

 

2. please add ref. for all equations and add latest ref. post 2020.

Thank you very much for your suggestion. A number of post-2020 references have been incorporated:

45. Gafni, T.; Shlezinger, N.; Cohen, K.; Eldar, Y.C.; Poor, H.V. Federated Learning: A Signal Processing Perspective. IEEE Signal Processing Magazine 2022, 39, 14–41, doi:10.1109/MSP.2021.3125282.
46. Hu, Z.; Shaloudegi, K.; Zhang, G.; Yu, Y. Federated Learning Meets Multi-Objective Optimization. IEEE Transactions on Network Science and Engineering 2022, 9, 2039–2051, doi:10.1109/TNSE.2022.3169117.
47. Yang, B.; Lei, Y.; Li, X.; Li, N.; Si, X.; Chen, C. A Dynamic Barycenter Bridging Network for Federated Transfer Fault Diagnosis in Machine Groups. Mechanical Systems and Signal Processing 2025, 230, 112605, doi:10.1016/j.ymssp.2025.112605.
48. Ruiz, E.; Bustos, A.; Rubio, H.; Castejón, C. Application of automatic classifiers for condition monitoring of railway rolling stock. Técnica Industrial, 2023, 336: 38-45. Doi: 10.23800/10542
49. Bustos, A.; Rubio, H.; Peláez, G.; García-Prada, J.C. Clasificación del estado de un eje ferroviario mediante EMD y clasificadores automáticos (2023). XXIV Congreso Nacional de Ingeniería Mecánica (CNIM-2023). Anales de Ingeniería Mecánica, año 23 (ISSN 0212-5072). In Spanish.

 

3. please put accelerometer specification as a table, what method was used to mount it on test bed?

Thank you very much for your suggestion. According to your recommendation, the specifications of the accelerometer are listed in a single table.

Parameter	Value
Sensitivity (±20%)	10.2mv(m/s2)
Acceleration range	± 490 m/s2
Frequency range (±3 dB)	.52 Hz to 8 kHz
Temperature range	-54º to 121ºC
Resonance frequency	25 kHz
Amplitude linearity	±1%
Transverse sensitivity	≤7%

 

The accelerometers are fixed with screws on the axle box of the bogie.

4. which type of chamber was used for test rig?? how you consider resonance??

Thank you very much for your interest. The test rig is installed inside the main workshop that RENFE company owns in the South of Madrid. RENFE is the major railway operator in Spain. Within this installation, a space is fitted out for bogie testing. This space is limited by four walls that prevent unauthorised people from accessing the test rig.

In this case, resonance is not considered due to the type of operation.

5. what is significance of depth e in line number 247??

Thank you for your hint. The parameter e is used to denote the depth of the defect or fault. It’s also shown in figure 3 which was improved as well:

 

6. please add a section about role of vibration as condition monitoring tool?

Thank you very much for your comment. According to your recommendation, a section was added.

1.2. Vibration analysis: condition monitoring tool

Vibration signal analysis is one of the most widely used techniques for inspecting mechanical components under operational conditions. [17], since the early models such as Mcfadden and Smith [18] in 1984 “Model for the vibration produced by a single point defect in a rolling element bearing”, as it allows for carrying out tests on a wide range of elements and situation [19], such as are railway infrastructure, general porpoise equip-ment or rolling-stock.

Many authors who study vibrations for condition monitoring apply their techniques to rolling elements, including Antoni and Randal [20] analysing bearing faults. Many of them have dealt with vibration analyses on railway systems [21], many focus on the ground or track perturbations, which are induced by the transit of rolling stock. This is as referred to by Bustos et al. 2018 [22].

Analysing vibration signals is a common testing ground for a wide range of me-chanical components Guo et al. [23] analyse signals in order to determine        faults on bearings, other authors male analysis for axle components as Luo et al. [24] 2020 or Bor-ghesani et al. in 2022 [25].

 

 

7. please add structure of paper at end of introduction section, how was figure 5 decomposition achieved??

Thank you very much for your hint. Authors consider it is a good idea to add structure of paper and according to your recomendation:

1.7. Structure of the paper

The structure of the paper is as follows: The following section details the methodolo-gy and techniques applied. Section 3 details the experimental system, test bench and sig-nal processing. Section 4 shows the results of model training. Section 5 contains the con-clusions of this work as well as further research suggestions.

Thank you very much for your comment. According to your suggestion the explanation provided in relation to the method by which Figure 5 was obtained has been enhanced.

During the process of EMD decomposing the signals into their IMF components, using MATLAB® software, it was found that the maximum number of components common to all signals was nine, although some signals reached up to ten components. In this study, the first nine IMFs of all signals were considered to homogenise the study.

Reviewer 3 Report

Comments and Suggestions for Authors

In response to safety accidents that may be caused by damage to railway vehicle bearings or axle components, this article proposes an automated fault diagnosis method based on EMD decomposition and machine learning, which has clear application scenarios and engineering significance. Some comments are listed as follows to improve the submission.

 

1. It is recommended to compare the decision tree method used with other machine learning methods to enhance the persuasiveness of the method selection.
2. The content of the Introduction is too scattered, and the background literature can be appropriately classified and integrated to avoid too much description. It is recommended to shorten the lengthy introduction of general methods such as AI and EMD and focus more on the research status and challenges in the field of railway vibration signal analysis.
3. It is suggested to provide a clear explanation of the selection process of experimental parameters, such as the number of IMFs, feature combinations, etc., rather than just listing the results after different combinations.
4. The confusion matrices such as Figure 6 and Figure 7 can be accompanied by specific category labels to enhance the intuitiveness of the results.
5. Data decentralization is an important issue in the field. The authors could discuss this by considering some works. -A dynamic barycenter bridging network for federated transfer fault diagnosis in machine groups.

Author Response

In response to safety accidents that may be caused by damage to railway vehicle bearings or axle components, this article proposes an automated fault diagnosis method based on EMD decomposition and machine learning, which has clear application scenarios and engineering significance. Some comments are listed as follows to improve the submission. 

We would like to express our sincere appreciation for their valuable comments, which have led us to modify the article accordingly. Thus, the improved version is not only the result of the authors' work, but also of the reviewers' hard work. All changes made as a result of the reviewers' comments are shown in blue in the revised article.

1. It is recommended to compare the decision tree method used with other machine learning methods to enhance the persuasiveness of the method selection.

Thank you for your hint. In the early stages of this project, several Machine Learning methodologies were employed, including K-nearest neighbour and Support Vector Machines. However, these methodologies yielded highly similar results, which ultimately led to the decision to adopt the methodology outlined in this study. The following references are added for clarification

48. Ruiz, E.; Bustos, A.; Rubio, H.; Castejón, C. Application of automatic classifiers for condition monitoring of railway rolling stock. Técnica Industrial, 2023, 336: 38-45. Doi: 10.23800/10542
49. Bustos, A.; Rubio, H.; Peláez, G.; García-Prada, J.C. Clasificación del estado de un eje ferroviario mediante EMD y clasificadores automáticos (2023). XXIV Congreso Nacional de Ingeniería Mecánica (CNIM-2023). Anales de Ingeniería Mecánica, año 23 (ISSN 0212-5072). In spanish.

And the following sentences:

The present work explores the use of decision trees combined with EMD decomposition as an alternative method for classifying and categorising the condition of railway rolling stock. This is achieved analysing the different stages of deterioration of the com-ponents, with the goal of determining the initial state of mechanical failure and improving general predictive maintenance, which offers results comparable to those obtained by Ruiz et al. [48], whose work was based on the use of Support Vector Machine, and Bustos et al. [49], who uses EMD and K-Nearest Neighbour to perform classification under simi-lar conditions.

2. The content of the Introduction is too scattered, and the background literature can be appropriately classified and integrated to avoid too much description. It is recommended to shorten the lengthy introduction of general methods such as AI and EMD and focus more on the research status and challenges in the field of railway vibration signal analysis.

Thank you very much for your hint. According to your recommendation the introduction has been reorganised. It is now divided into several sections, including Approaches and challenges and Objective of the study. We have made some modifications in order to be in line with your suggestions.

.

3. It is suggested to provide a clear explanation of the selection process of experimental parameters, such as the number of IMFs, feature combinations, etc., rather than just listing the results after different combinations.

Thank you very much for your appreciation. According to your recommendation the explanation of the training method has been improved to make the methodology used to obtain the results clearer.

1. Model is trained for every hyperparameter. Obtained results are registered.
2. The hyperparameters are then combined in groups of two, three and so on, until all hyperparameters have been combined for each class, time and frequency, and at each end of the axle.
3. The distinct combinations of the hyperparameters belonging to the two classes, time and frequency, are achieved through the variation of the composition of the constituent groups of hyperparameters undergoing training. This process is pursued until all the hyperparameters of the two classes are utilised for each end of the axle.
4. Steps 1 to 3 are performed by taking all the hyperparameters for the axis as a whole without distinguishing the ends.

Following the generation of each training-related Confusion Matrix and ROC curve, subsequent analytical procedures may be initiated.

In this work, we have considered the combinations of hyperparameters shown are selected so that have yielded significant outcomes. These combinations are presented for the purpose of comparison, with the aim of assessing the efficacy of the method as a whole.

 

4. The confusion matrices such as Figure 6 and Figure 7 can be accompanied by specific category labels to enhance the intuitiveness of the results.

Thank you very much for your hint. According to your recommendation, a reminder was added before figures:

As it was previously commented in section 3.2, four conditions are defined: undamaged Axle (D0, 0.0 mm), damaged Axle Defect 1 (D1, 5.7 mm), damaged Axle Defect 2 (D2, 10.9 mm) and damaged Axle Defect 3 (D3, 15.0 mm).

Figures descriptions are accordingly modified.

5. Data decentralization is an important issue in the field. The authors could discuss this by considering some works. -A dynamic barycenter bridging network for federated transfer fault diagnosis in machine groups.

Thank you very much for your appreciation. According to your recommendation, a subsection entitled 'Approaches and challenges' has been added to the introduction. This provides a reference point for various developments using other methodologies, as well as the challenges facing the field.

About other approaches that have been taken to achieve the goal of optimising the analysis of incoming data, Gafni et al. [45] identify Federated Learning as a tool for the future to, among other things, use information received from different sources which, in many cases, may not be fully reliable and/or relevant, to a certain extent, Hu et al. [46] are associated with not only different data sources, but also different research aims and the improvement of data. The amount of data gives rise to a number of challenges, but it also creates an area of interest in the form of decentralised data, as has been pointed out by Yang et al. [47], but the quality of the information, and thus the effectiveness of the analysis, is affected by the use of data from different sources and the methods of data collection.

45. Gafni, T.; Shlezinger, N.; Cohen, K.; Eldar, Y.C.; Poor, H.V. Federated Learning: A Signal Processing Perspective. IEEE Signal Processing Magazine 2022, 39, 14–41, doi:10.1109/MSP.2021.3125282.
46. Hu, Z.; Shaloudegi, K.; Zhang, G.; Yu, Y. Federated Learning Meets Multi-Objective Optimization. IEEE Transactions on Network Science and Engineering 2022, 9, 2039–2051, doi:10.1109/TNSE.2022.3169117.
47. Yang, B.; Lei, Y.; Li, X.; Li, N.; Si, X.; Chen, C. A Dynamic Barycenter Bridging Network for Federated Transfer Fault Diagnosis in Machine Groups. Mechanical Systems and Signal Processing 2025, 230, 112605, doi:10.1016/j.ymssp.2025.112605.

Reviewer 4 Report

Comments and Suggestions for Authors

The study presents a method for assessing the condition of rolling stock, particularly damage to wheel sets, based on the measurement of vibroacoustic signals and the application of machine learning techniques. The entire study is consistent with its title and meets the current needs for forecasting the wear and damage of rail vehicle components that affect rail transport safety.
The paper presents the advantages of machine learning for processing large amounts of data collected during train movement while monitoring physical parameters and vibrations in selected locations of a rail vehicle. The use of Empirical Mode Decomposition (EMD) was necessary for the results presented in the study. The process of using EMD for diagnostics is illustrated in Fig. 1, specifically for the diagnostics of rolling bearings in wind turbines. In the introduction to the study, it was essential to analyze the machine learning methods used for diagnosing mechanical devices. Based on an analysis of available machine learning techniques for vibration signal processing, the authors propose their methodology for vibration signal processing. The entire second chapter of the thesis is devoted to a detailed description of the proposed method for processing vibration signals. The authors suggested using MATLAB software to read and process signals obtained from experimental tests. They described in detail the signal criteria checked by the algorithm prepared in MATLAB. They presented a 7-point algorithm for extracting the signal of each internal modal function (IMF) found in the analyzed measurement results.
Chapter 3, presenting the research stand, was equally important in the thesis. The authors used bench tests of a Y-21 freight wagon. The test results concerned the vibrations of a wheelset with a geometric defect (loss) of the wheelset axle. The time course of the wheelset accelerations is shown in Fig. 4. Using the vibration signal, the authors present its decomposition into nine internal modal functions and demonstrate how these can be used to train and test the model.
When summarising the model results in Chapter 4, the authors emphasized the achievement of high results in most cases, exceeding 90%. A detailed analysis of the results' impact was presented by the authors in the final chapter, along with their conclusions.
The authors have presented the logic and methodology of the proposed actions with great care. The work addresses the critical issue of improving the safety and reliability of rail vehicle operation and, thanks to the use of actual measurements from bench tests of a bogie, constitutes a valuable and comprehensive source of information.
The authors made a mistake in the numbering of the chapters; instead of the number 5, there is the number 6. A substantive inaccuracy in the article is the inclusion of 'high-speed trains' in the keywords, whereas the research experiment refers to a freight wagon bogie. I suggest that the authors consider how to alleviate this inconvenience (perhaps by modifying the keyword).
Most of the drawings are clear; only drawing no. 5 could have larger areas showing the vibration patterns at the expense of the axis descriptions.
The conclusions presented are comprehensive, and the illustrative material and literature sources have been selected appropriately. The authors have cited 44 literature sources, including both recent ones from the last few years and older ones dating back over 20 years. Each of the sources indicated is relevant to the proper context of the work. Most of the figures are clear; however, Figure 5 could be improved by allocating larger areas to the vibration patterns at the expense of the axis descriptions.

Author Response

The study presents a method for assessing the condition of rolling stock, particularly damage to wheel sets, based on the measurement of vibroacoustic signals and the application of machine learning techniques. The entire study is consistent with its title and meets the current needs for forecasting the wear and damage of rail vehicle components that affect rail transport safety.

The paper presents the advantages of machine learning for processing large amounts of data collected during train movement while monitoring physical parameters and vibrations in selected locations of a rail vehicle. The use of Empirical Mode Decomposition (EMD) was necessary for the results presented in the study. The process of using EMD for diagnostics is illustrated in Fig. 1, specifically for the diagnostics of rolling bearings in wind turbines. In the introduction to the study, it was essential to analyze the machine learning methods used for diagnosing mechanical devices. Based on an analysis of available machine learning techniques for vibration signal processing, the authors propose their methodology for vibration signal processing. The entire second chapter of the thesis is devoted to a detailed description of the proposed method for processing vibration signals. The authors suggested using MATLAB software to read and process signals obtained from experimental tests. They described in detail the signal criteria checked by the algorithm prepared in MATLAB. They presented a 7-point algorithm for extracting the signal of each internal modal function (IMF) found in the analyzed measurement results.

Chapter 3, presenting the research stand, was equally important in the thesis. The authors used bench tests of a Y-21 freight wagon. The test results concerned the vibrations of a wheelset with a geometric defect (loss) of the wheelset axle. The time course of the wheelset accelerations is shown in Fig. 4. Using the vibration signal, the authors present its decomposition into nine internal modal functions and demonstrate how these can be used to train and test the model.

When summarising the model results in Chapter 4, the authors emphasized the achievement of high results in most cases, exceeding 90%. A detailed analysis of the results' impact was presented by the authors in the final chapter, along with their conclusions.

The authors have presented the logic and methodology of the proposed actions with great care. The work addresses the critical issue of improving the safety and reliability of rail vehicle operation and, thanks to the use of actual measurements from bench tests of a bogie, constitutes a valuable and comprehensive source of information.

We would like to express our sincere appreciation for their valuable comments, which have led us to modify the article accordingly. Thus, the improved version is not only the result of the authors' work, but also of the reviewers' hard work. All changes made as a result of the reviewers' comments are shown in blue in the revised article.

From a more personal point of view, the effort, dedication, appreciation, recognition and hard work analysing this project shown by someone with such extensive experience and knowledge is a great encouragement for future projects.

1. The authors made a mistake in the numbering of the chapters; instead of the number 5, there is the number 6.

Thank you for your hint. According to your recommendation the mistake has been fixed

2. A substantive inaccuracy in the article is the inclusion of 'high-speed trains' in the keywords, whereas the research experiment refers to a freight wagon bogie. I suggest that the authors consider how to alleviate this inconvenience (perhaps by modifying the keyword).

Thank you for your hint. Keyword was changed to freight train

3. Most of the drawings are clear; only drawing no. 5 could have larger areas showing the vibration patterns at the expense of the axis descriptions. The conclusions presented are comprehensive, and the illustrative material and literature sources have been selected appropriately. The authors have cited 44 literature sources, including both recent ones from the last few years and older ones dating back over 20 years. Each of the sources indicated is relevant to the proper context of the work. Most of the figures are clear; however, Figure 5 could be improved by allocating larger areas to the vibration patterns at the expense of the axis descriptions.

Thank you very much for your observation. According to your recommendation, Figure 5 has been improved. It is a good idea, the descriptions have been removed as they are redundant and unnecessary for the purpose of the figure, which is, as you rightly point out, to illustrate the different signals by enlarging the area in which they appear:

 

Round 2

Reviewer 3 Report

Comments and Suggestions for Authors

The authors have well addressed all my comments, and it can be accepted now.