Gear Fault Diagnosis through Vibration and Acoustic Signal Combination Based on Convolutional Neural Network
Round 1
Reviewer 1 Report
- Introduction, line 28, when you mention eliminated? What does it mean? Since the monitoring and diagnostic systems do not be able to eliminate machine faults, it just detects and identifies the mechanism that suffers it.
- In the dataset, why the signals at 1800 and 2700 rpm are not used? It could complement the analysis.
- In the data set, the vibration sensor is placed in a vertical, horizontal, or axial position? It influences the obtained result.
- What previous study was used to establish the wavelet parameters calculation? Or what technique was used to fix those parameters? It applies to vibration signals characterization.
- It is widely known that gearbox signals and the studied faults at constant speed are considered stationary signals. So, a comparison with the time series or frequency domain is mandatory.
- Inline 262, and Figure 11. It is mentioned comparison with methods like FFT-SVM and FFT-MLP, but those methods are not based on deep learning, and neither the references [18] and [19] are related to them. In contrast, [18] uses a Deep Neural Network (DNN) and [19] Convolutional Neural Network with Wide first-layer kernels (WDCNN), to diagnostic bearing faults. Both are considered state-of-the-art methods in deep learning classification. Therefore, include testing with both approaches is needed. (It is recommended to sound signal analysis also.)
- Evidently, in the classification process, the feature sets used are not similar, and hence, it influences the obtained performance for all classifiers. To performance comparison, it is required to use the same feature set.
- The parameters of MLP and SVM are not explained either mentioned. Some similar occurs in sound analysis with ANN classifier.
Author Response
Original Manuscript ID: information-772798
Original Article Title: “Gear Fault Diagnosis through Vibration and Acoustic Signal Combination based on Convolutional Neural Network”
To: Information Editor
Re: Response to reviewers
Dear Editor/Reviewer,
Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments.
We are uploading (a) our point-by-point response to the comments (below) (response to reviewers), (b) We re-polished the language of this paper, and (d) an updated manuscript with red highlighting indicating changes.
We tried our best to improve the manuscript and made substantial changes in the introduction.
Once again, thank you very much for your comments and suggestions. We appreciate for Editors/Reviewers’ warm work earnestly and hope that the correction will meet with approval.
Best regards,
Liya Yu.
Reviewer 1：
- Concern # 1:Introduction, line 28, when you mention eliminated? What does it mean? Since the monitoring and diagnostic systems do not be able to eliminate machine faults, it just detects and identifies the mechanism that suffers it.
Answer: Thank you for your suggestions. I agree with your view, the monitoring and diagnostic systems do not be able to eliminate machine faults, it just analyzes the collected data and gives the failure rate of the corresponding equipment. Technicians decide what maintenance strategy to take by looking at the analysis results. Only by mastering the health status of the equipment in time, can the hidden trouble be found and eliminated effectively by the technicians, not by diagnostic systems automatically, so the word “eliminated” is to described for hidden troubles, this is my initial meaning.
Author action: In order to avoid misunderstanding of the meaning, we replace the sentence as the following and marked with red in the text:
Only by mastering the health status of the equipment in time, can the hidden trouble be found and eliminated effectively by technicians.
- Concern # 2:In the dataset, why the signals at 1800 and 2700 rpm are not used? It could complement the analysis.
Answer: Thank you for your suggestion. Our dataset, as shown in Table 1, a total of 10 different operating conditions that correspond to the three speeds (900r/m, 1800r/m, 2700r/m) of the motor were simulated to ensure the diversity of the samples. Experimental data are obtained from the raw data by random sampling to objectively evaluate the performance of the proposed model. The vibration dataset contains 200 samples for each type of gear fault. The acoustic dataset was similarly created.
In 3.1 Feature Extraction, because the time-frequency graphs at different speeds are similar, in order to introduce this part briefly, we just present time domain, spectrum and time–frequency diagrams of normal gears, worn gears, broken gears and pitting gears under the speed of 900r/m, which are shown in Fig. 4-7. And in the following parts, the experiments analysis is based on the all speeds condition data (900r/m, 1800r/m, 2700r/m).
- Concern # 3:In the data set, the vibration sensor is placed in a vertical, horizontal, or axial position? It influences the obtained result.
Answer: Thank you for your suggestion and reminding, we will add this detail to the content. We adopt horizontal installation of vibration sensor during the experiment to collect data, this installation has been widely used in the existing literature and similar experiments. Thanks to the expert ’s suggestion. In the later research, we will try to see if different installation methods will have an impact on fault diagnosis results.
Author action: We add the detail of the vibration sensor installation to the Experimental setup content and marked with red in the text.
The CY1010L piezoelectric accelerometer is mounted horizontally on the side of the gearbox for vibration signals.
- Concern # 4:What previous study was used to establish the wavelet parameters calculation? Or what technique was used to fix those parameters? It applies to vibration signals characterization.
Answer: Thank you for your suggestion. Wavelet transform is only a tool used in this section. With the help of it, the time-frequency image of signal can be obtained, which is connected with CNN of image processing. Therefore, the selection and modification of wavelet parameters are not discussed, and the empirical values of other literatures are directly applied.
- Concern # 5:It is widely known that gearbox signals and the studied faults at constant speed are considered stationary signals. So, a comparison with the time series or frequency domain is mandatory.
Answer: Thank you for your suggestion. In this paper, we compare and analyze the time domain, spectrum and time-frequency of the signal, which are described in Section 3.1.
- Concern # 6:Inline 262, and Figure 11. It is mentioned comparison with methods like FFT-SVM and FFT-MLP, but those methods are not based on deep learning, and neither the references [18] and [19] are related to them. In contrast, [18] uses a Deep Neural Network (DNN) and [19] Convolutional Neural Network with Wide first-layer kernels (WDCNN), to diagnostic bearing faults. Both are considered state-of-the-art methods in deep learning classification. Therefore, include testing with both approaches is needed. (It is recommended to sound signal analysis also.)
Answer: Thank you for your suggestions. The purpose of citing these two literatures here is that they are all based on Case Western Reserve University's bearing fault data for diagnostic analysis, and [19] use the fft-svm and fft-mlp methods as the compared methods. Based on existing literature, fft-svm and fft-mlp are common machine learning methods in the analysis and processing of vibration signals. Therefore, the method designed in this paper are compared with them to highlight the effectiveness of our proposed methods.
- Concern # 7: Evidently, in the classification process, the feature sets used are not similar, and hence, it influences the obtained performance for all classifiers. To performance comparison, it is required to use the same feature set.
Answer: Thank you for your suggestions. The feature sets influence the obtained performance of classifier. For the same signal, when using different models to classify it, the feature set we adopt is consistent.
- Concern # 8:The parameters of MLP and SVM are not explained either mentioned. Some similar occurs in sound analysis with ANN classifier.
Answer: Thank you for your suggestions. In the analysis of vibration signal, MLP and SVM is introduced to highlight the effectiveness of ASCNN, so this paper focuses on the ASCNN, including model composition, parameter setting and performance analysis. Similarly, ANN is introduced to highlight ESCNN for comparison in the analysis of sound signals, so this paper focuses on the ESCNN.
Author Response File: 
Author Response.docx
Reviewer 2 Report
The article entitled "Gear Fault Diagnosis through Vibration and Acoustic Signal Combination based on Convolutional Neural Network" evaluates a gearbox fault diagnosis based on vibration and sound signals. The authors used a convolutional neural network as a proposed method.
Writing & Presentation
There are formatting mistakes in the text, which requires a review. Some are listed with additional comments following:
- Review the topics numbering. Introduction should starts from #1;
- Axis of figures 4 to 7 misses the amplitude unit;
- Page 3, line 114 – Review the reference formatting from [13];
- Page 8, line 190 – Review the letter case in the beginning of the phrase;
- There are several acronym terms without indication of the meaning. For example, MNIST, IDS, IDSCNN etc.
Technical content
* The paper misses a more detailed theoretical description that was used in the methods. Also, there is no reference about the implemented algorithm. Which software was used in the work?
* In the real gearbox, there are different noise sources. How that algorithm would work within this condition? Will it be effective to identify the failure cause?
* The time-frequency graphics, Figures 4 to 7, should be rescaled in order to view the maximum amplitude in the map (the red color is not visible).
Author Response
Original Manuscript ID: information-772798
Original Article Title: “Gear Fault Diagnosis through Vibration and Acoustic Signal Combination based on Convolutional Neural Network”
To: Information Editor
Re: Response to reviewers
Dear Editor/Reviewer,
Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments.
We are uploading (a) our point-by-point response to the comments (below) (response to reviewers), (b) We re-polished the language of this paper, and (d) an updated manuscript with red highlighting indicating changes.
We tried our best to improve the manuscript and made substantial changes in the introduction.
Once again, thank you very much for your comments and suggestions. We appreciate for Editors/Reviewers’ warm work earnestly and hope that the correction will meet with approval.
Best regards,
Liya Yu.
Reviewer3:
Concern # 1:the article entitled "Gear Fault Diagnosis through Vibration and Acoustic Signal Combination based on Convolutional Neural Network" evaluates a gearbox fault diagnosis based on vibration and sound signals. The authors used a convolutional neural network as a proposed method.There are formatting mistakes in the text, which requires a review. Some are listed with additional comments following:
- Review the topics numbering. Introduction should starts from #1;
- Axis of figures 4 to 7 misses the amplitude unit;
- Page 3, line 114 – Review the reference formatting from [13];
- Page 8, line 190 – Review the letter case in the beginning of the phrase;
- There are several acronym terms without indication of the meaning. For example, MNIST, IDS, IDSCNN etc.
Answer: Thank you for your suggestions and sincerity. As for the mistakes in writing and presentation which you pointed out, we have corrected and marked with yellow in the text. As for the fifth problem, we have added the corresponding meaning to the several acronym terms where they first appear. Thank you again for your suggestion, which inspires us to be more careful and serious in our research.
Concern # 2:* The paper misses a more detailed theoretical description that was used in the methods. Also, there is no reference about the implemented algorithm. Which software was used in the work?
Answer: Thank you for your suggestions. The more detailed theoretical description used in the method is not introduced in this paper, this paper is based on our previous research, the article of literature 13, which has been explained in the paper. We have marked it in yellow. At the same time, we used MATLAB and Python in our work. Thank you again for your suggestion.
Concern # 3:* In the real gearbox, there are different noise sources. How that algorithm would work within this condition? Will it be effective to identify the failure cause?
Answer: Thank you for your suggestions. I think your opinion is very worthy of study in later research. In practice, there are different noise sources that will affect the algorithm. How to deal with the problem of external noise is our next step. But in the experiment of this paper, we mainly focus on the effectiveness of our proposed method under an ideal environment, so we put the experimental platform in a semi anechoic environment, which can reduce the interference of external noise to a certain extent. Future research will be based on this method to explore the migration to actual noisy working conditions, which is also our next research direction, so this paper does not involve this part. So hereby explain to you clearly, thank you very much.
Concern # 4:* The time-frequency graphics, Figures 4 to 7, should be rescaled in order to view the maximum amplitude in the map (the red color is not visible).
Answer: Thank you for your suggestions. I am so sorry that the amplitude is too small to see the red maximum amplitude in Figure 4-7. In the normal state, the amplitude of the signal is concentrated below 200, in the worn state between 0-300, in the broken tooth state below 200, and in the pitting state below 200. I used MATLAB's toolbox for drawing, because the ordinate of the software is at 1000hz intervals, I looked for a lot of information, but I cannot modify the parameters, so the red maximum amplitude is not visible. Sorry for this, hope you can understand, thank you again.
Author Response File: Author Response.docx
Reviewer 3 Report
The authors propose a fault diagnosis scheme with a combination of acoustic signals and vibration based on the Convolutional Neural Network (CNN). They did a good experimental job, combining the most widely used sensor signals in the industry. Some references are missing in the introduction:
10.1109/TCSII.2019.2920609
https://0-doi-org.brum.beds.ac.uk/10.3390/mca19010037
https://0-doi-org.brum.beds.ac.uk/10.3390/mca22020030
https://0-doi-org.brum.beds.ac.uk/10.3390/pr7110814
In the review of state of the art on page 2, lines 54-72, the use of CNN is generally addressed. However, it is improved by citing papers where CNN is used in Gear Fault Diagnosis; In recent years, this type of neuronal network has been used for this purpose. Then, the contribution of the work is focused on the combination of two CNNs (ASCNN and ESCNN), later it is processed with the IDS theory. This last step should be highlighted in the section.
The experimental configuration is well described; why they adjusted the motor speeds to (900,1800 and 2700 r / m)?, What happens in the transition of those motor revolutions?
Finally, the performance analysis of the different methods is interesting; the effectiveness of the proposed method is demonstrated. Nevertheless, it remains to address how the performance of the method would be in the presence of external noise. For the reviewed version, it is necessary to consider noise.
Author Response
Original Manuscript ID: information-772798
Original Article Title: “Gear Fault Diagnosis through Vibration and Acoustic Signal Combination based on Convolutional Neural Network”
To: Information Editor
Re: Response to reviewers
Dear Editor/Reviewer,
Thank you for allowing a resubmission of our manuscript, with an opportunity to address the reviewers’ comments.
We are uploading (a) our point-by-point response to the comments (below) (response to reviewers), (b) We re-polished the language of this paper, and (d) an updated manuscript with red highlighting indicating changes.
We tried our best to improve the manuscript and made substantial changes in the introduction.
Once again, thank you very much for your comments and suggestions. We appreciate for Editors/Reviewers’ warm work earnestly and hope that the correction will meet with approval.
Best regards,
Liya Yu.
Reviewer2：
Concern # 1:The authors propose a fault diagnosis scheme with a combination of acoustic signals and vibration based on the Convolutional Neural Network (CNN). They did a good experimental job, combining the most widely used sensor signals in the industry. Some references are missing in the introduction:
Answer: Thank you for your suggestions. I think your recommended literatures are very useful, and we have added it to the introduction.
Author action: We have added the above recommended literatures in the introduction and update the whole references, and marked with yellow in the text. Thanks again for your recommendation.
- Martínez-García C, Astorga-Zaragoza C, Puig V, et al. A simple nonlinear observer for state and unknown input estimation: DC motor applications[J]. IEEE Transactions on Circuits and Systems II: Express Briefs, 2019.
- Yuzukirmizi M, Arslan H. Fault diagnosis of shaft-ball bearing system using one-way analysis of variance[J]. Mathematical and Computational Applications, 2014, 19(1): 37-49.
- López-Estrada F R, Rotondo D, Valencia-Palomo G. A Review of Convex Approaches for Control, Observation and Safety of Linear Parameter Varying and Takagi-Sugeno Systems[J]. Processes, 2019, 7(11): 814.
- Granda D, Aguilar W G, Arcos-Aviles D, et al. Broken bar diagnosis for squirrel cage induction motors using frequency analysis based on MCSA and continuous wavelet transform[J]. Mathematical and Computational Applications, 2017, 22(2): 30.
Concern #2:In the review of state of the art on page 2, lines 54-72, the use of CNN is generally addressed. However, it is improved by citing papers where CNN is used in Gear Fault Diagnosis; In recent years, this type of neuronal network has been used for this purpose. Then, the contribution of the work is focused on the combination of two CNNs (ASCNN and ESCNN), later it is processed with the IDS theory. This last step should be highlighted in the section.
Answer: Thank you for your suggestions. This part belongs to the introduction of the paper, which aims to give a general overview of the work done and the background. This paper focuses on the combination of ASCNN and ESCNN, which are based on CNN, so we make a more detailed introduction to CNN at first. The details of our proposed method are described in Section 5. To avoid repetition, details are not emphasized in the introduction. Thank you again for your valuable comments.
Concern #3: The experimental configuration is well described; why they adjusted the motor speeds to (900,1800 and 2700 r / m)? What happens in the transition of those motor revolutions?
Answer: Thank you for your suggestions. Because related research has shown that machine fault under different motor speeds would have different characteristics (such as amplitude, frequency and so on), the similar experimental configuration was also applied in Case Western Reserve University's bearing fault data collection. When we adjust the motor speed to 900 r /m,1800 r /m and 2700 r /m, we could study the gear fault under more comprehensive working conditions, to further prove the effectiveness of our proposed method as well.
Author action: We add the detail of the purpose for the motor speeds in 2 Experimental Setup to make readers more clearly, and marked with red in the text.
The motor speed was adjusted to 900, 1800, and 2700 r/m to simulate under different working conditions.
Concern # 4: Finally, the performance analysis of the different methods is interesting; the effectiveness of the proposed method is demonstrated. Nevertheless, it remains to address how the performance of the method would be in the presence of external noise. For the reviewed version, it is necessary to consider noise.
Answer：Thank you for your suggestions. I think your opinion is very worthy of study in later research. The external noise does exist in real workshop. But in the experiment of this paper, we mainly focus on the effectiveness of our proposed method under an ideal environment, so we put the experimental platform in a semi anechoic environment, which can reduce the interference of external noise to a certain extent. Future research will be based on this method to explore the migration to actual noisy working conditions, which is also our next research direction, so this paper does not involve this part. So hereby explain to you clearly, thank you very much.
Author Response File: Author Response.docx
Round 2
Reviewer 2 Report
The paper supposed to present the main theory, with a summary if it was presented in a previous reference. This paper still misses the structure of the scientific work.
Reviewer 3 Report
Authors have addressed all my concerns. It can be accepted in the present form.
