Condition-Based Maintenance for Normal Behaviour Characterisation of Railway Car-Body Acceleration Applying Neural Networks

Round 1

Reviewer 1 Report

The paper presents a study regarding the application of neural network algorithms to deal with the condition-based maintenance (CBM). This is an interesting subject that deserves to be studied. However, before publication, the following comments should be considered and answered to improve the quality of the work.

 

1 – General: the main abbreviation used in this paper (CBM - Condition based maintenance?) is never defined. All the abbreviations must be defined in the first time they appear in the paper.

 

2 – Introduction: reference [1] is defined in the reference list as “DIN EN 14363”. The full description of the norm, including year, publisher, title, has to be defined. Make sure all the citations are correctly defined in the list.

 

3 – Introduction: the European norm dedicated to measuring comfort should be cited (EN12299 – 2009: Railway applications - Ride comfort for passengers - Measurement and evaluation).

 

4 – Introduction: A final paragraph should be added to the Introduction clearly stating the main innovations of the paper.

 

5 – Lines 117-126: in the definition of the variables of an equation, the definition should be given in the same sentence, ie, “where z represents…, wi represents…etc”. Writing “where” all the time in different paragraphs is not well written. Please, revise all document and change accordingly. Moreover, all sentences from Section 2 are in different paragraphs, which is not also corrected. Please revise the whole document to make it more readable (same ideas in the same paragraph, do not need to change paragraph all the time).

 

6 – Section 3: a more detailed information regarding the experimental setup carried out in 2017 should be given. The authors say also in Section 4 that a dataset obtained during a whole year was used, but nothing is said regarding the test setups (type of sensors, acquisition rates, train route, etc). Special attention should be given to the acquisition system used. The reviewer is curious about this, since there are dozens of sensors throughout all coaches, and it would be interesting to understand how the sensors were synchronized.

 

7 – Figure 5: The axle’s legends and values are not readable (too small). Moreover, then this figure is quoted in the text it says “As shown in Figure 6, dynamic variables (speed and accelerations) are plotted close 206 to each other”, however it should be Fig. 5. Check the whole document and correct this in the remaining figures. Apart from these issues, a more technical discussion regarding this graphic should be given. The authors say that they perform a PCA to obtain these correlations, but a more clear/detailed explanation should be given in this regard.

 

8 – Figure 6: Why the accelerometers were positioned in the upper part of the coach’s endwall? Usually they are placed in the floor or seats for comfort evaluations.

 

9 – Lines 265-266: Why the target variable is the lateral acceleration of the carbody and not the other ones (ayr, azc)?

 

10 – Lines 322-324: Could the authors explain this sentence: “One second is not enough for an acceleration or speed value to change, so consecutive samples measuring the same acceleration value are regularly found in the dataset.”. In a normal sensor with acquisition rate of 2kHz there are 2000 readings in each second, therefore, the Reviewer did not understand this sentence.

 

11 – Lines 358-359: Could the authors explain this sentence: “In addition, the number of samples predicted with a mean 358 absolute error less than 0,25 m/s2 is 0% in nearly every case.”. Errors LESS than 0.25m/s2 are 0%? The mean absolute error column shows errors less than 0.1m/s2, not higher than 0.25ms/2.

 

12 – Figure 9: Could the authors explain the big differences between measured and predicted values of the lateral carbody accelerations in axle 11. Moreover, please add axle titles to the vertical axes of all figures. The size of the numbers is small, and the figures are with bad quality.

 

13 – Section 6: this section is quite vague without any results, since it just gives a description of how the methodology can be generalized to other situations. Therefore, it is the Reviewer’s opinion that this should not be a isolated section and it should be incorporated in the previous one.

Author Response

Dear reviewer,

Please see the attachment with the answer to your comments.

Thank you very much and best regards,

Prof. Pablo Garrido

Author Response File: Author Response.pdf

Reviewer 2 Report

This manuscript presents experimental research of the solution for condition-based monitoring of the rolling stock. The research is important for practical engineering and railway operators. The manuscript presents a lot of standard information about the methods but not much specific technical information about the measurement process (sensors used and their locations). The aim of the paper and the sense of the “prediction” is not fully clear. Evidently, the prediction of the bogie lateral accelerations is in the time horizon 400s is meant. However, it is no clear how it corresponds to the maintenance operation. Thus, it is questionable if there is a use of such prediction of the accelerations. The general concept is not clear – the data were used form a normal behaviour of railway carbody and the not normal behaviour is predicted. The manuscript should be substantially revised.

The specific comments:

• The literature review should be in improved with the new relevant literature and the reference list too short.
• Fig.3-4 Y- axis title is missing.
• Fig.5 – improve the font and append the PC score points. The PCA in the present form with the data set from the normal state does not bring any new conclusions. Minimal two states should be considered – normal and failure.
• Measurement rate 1 sec for the high-frequency oscillations?
• What reasons (failure modes) could cause the comfort degradation? How could be decided that there is a problem in the track or the rolling stock or both of them?
• The conclusion in lines 210-214 was done for one normal mode of operation, therefore, there is no relation to the temperature. However, in the failure mode – it could have a high influence.
• More detailed locations of the sensors in Fig. 6 should be presented and photos if available.
• What are predictors variables and a target variable in the dataset?
• Fig.9 presents an increase of the lateral carbody accelerations (blue line) but the model predicts no increase of the acceleration and “a clear comfort degradation” is concluded. The conclusion is not clear. What is the use of the acceleration prediction, if it is measured at the same time?

Author Response

Dear reviewer,

Please see the attachment with the answer to your comments.

Thank you very much and best regards,

Prof. Pablo Garrido

Author Response File: Author Response.pdf

Reviewer 3 Report

1) In the introduction, generally, the reviewer thinks that the literature review is not quite sufficient. Actually, a number of papers focusing on carbody acceleration are not involved. 15 references seem to be not enough for an academic article. Firstly, it is recommended to indicate that the violent vibration of the carbody is normally caused by the track irregularities [*1] and wheel-rail perturbations [*2]. Secondly, it is suggested that some literatures should be included to highlight the motivation of predicting the carbody acceleration. What exactly is the carbody acceleration used for?

[*1] “A spatial coupling model to study dynamic performance of pantograph-catenary with vehicle-track excitation,” Mech. Syst. Signal Process., vol. 151, p. 107336, 2021.

[*2] "Fundamentals of vehicle–track coupled dynamics." Vehicle System Dynamics 47.11 (2009): 1349-1376.

 

2) In Section 2, the labels on y-axis are missing for figures 3-5.

 

3) Please specify what are exactly the inputs and outputs for the neural network.

 

4) Again, the y-axis labels are missing for Figure 8-9.

 

5) In Figure 9, it looks like that the predicted acceleration has a big difference from the real acceleration after 200 s. How do you claim the validation of the present model?

 

6) The texts in Figure 2 are totally unreadable. Please consider improving.

 

7) Please give more details about how the acceleration data are collected. For instance, how much is the train speed?

 

8) It is also recommended to plot the structure of the neural network to indicate how many layers are used in the ANN.

Author Response

Dear reviewer,

Please see the attachment with the answer to your comments.

Thank you very much and best regards,

Prof. Pablo Garrido

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

the paper can be accepted in its current form

Author Response

Thank you very much for your professional review

Reviewer 2 Report

This manuscript text is substantially improved. However, the intrinsic problem was not removed. Now, the author state that the lateral car-body accelerations are predicted by all other ones at the same time moment. Such prediction has low sense and is not mechanically substantiated. However, a prediction of future accelerations would be very plausible. I suppose the authors aim to predict rapid deterioration of the vehicle mechanical parts based on the measured data in the last history of the trains. However, in this case, the prediction horizon should be defined: N days/weeks/months to the expected deterioration. The information would be relevant to the maintenance operators. The model learning should be based on such cases of deterioration. I believe the available measurement dataset from the long period of time 2005-2017 should include enough cases of rapid deterioration.

Not all comments were fully answered:

• Fig.7 is too simple – PC points would improve it. Please see the studies with PCA visualization, especially for railway crossing condition monitoring.
• The conclusion the there is no relation of the acceleration to the temperature is done because the same time moment is considered. However, temperature is an important indicator of future mechanical deterioration.
• “More detailed location of the sensors in Fig. 6 should be presented and photos if available.” Fig. 5 or 6 is not improved. It should include the main parts of the rail vehicle (box, bogie…) and the sensor position in it, sensor axis, etc. The photos are not added.
• The references [20] and [21] have no authors.

Author Response

Thank you very much. Please find attached the answers to your comments.

Author Response File: Author Response.docx

Reviewer 3 Report

Generally, I think this paper can be published with a good contribution to the community. Two minor editorial errors can be tackled in the publication process.

1) please reformat the references.

2) 'Contribute to' is followed by a noun phrase, not a verb. 

'Contribute to create' should be 'creating'.

Author Response

Thank you very much for your professional review