An Improved Fault Diagnosis Using 1D-Convolutional Neural Network Model

Round 1

Reviewer 1 Report

The paper describes a method for faults classification in roller bearings using Convolutional Neural Networks (CNN). The method is tested using the CWRU dataset and the architecture of the CNN is adapted for handling a 1-D input signal.

The research topic is important, however there are several research papers reporting CNN based classification methods using the CWRU dataset with accuracies of 100% or very close to 100%. The classification problem for this dataset has been solved using either 1-D or 2D CNN, attaining high accuracies.

 

The paper has several issues that must be improved:

1) The original contribution of this paper is rather marginal. Authors do not show outstanding improvements with respect to research reported in the literature using this dataset.

2) The methodology used is flawed as a cross-validation is not performed for the CNN model proposed and the other classification models used for comparison.

 

3) The written grammar in some paragraphs is erroneous. In consequence, it is difficult to follow the idea expressed by the authors.

4) In section (3, line 175) authors claim as a shortcoming for using 2D-CNN: “(1) a long time to transform the 1D vibration signals to 2D images”, however, this can be done using several approaches where the simpler is performed by reshaping an array and it actually do not take a long time.

5) Concerning the CNN architecture shown in Fig. 3, the number of filters and the size of the filters decreases along the CNN architecture. This is not the common structure for CNNs where the number of filters increases as the size of the filter decreases along the CNN network. Authors must justify this choice.

6) Authors claim in line 272 the realization of five training sessions, however this procedure does not correspond to a formal cross-validation of this approach.

7) The review presented in the introduction section concerning the CNN applications for roller bearings is rather poor.

Author Response

Thanks for the reviewer's suggests.

Each comment has been addressed carefully in our revised manuscript and the modifications are highlighted in red.

Author Response File: Author Response.pdf

Reviewer 2 Report

Dear Authors, I have read your submission with great interest.
I'm pleased to see that you have done quite extensive research.

In following numbered section I will try express my view of your work.

0. Title: And Improved of Fault Diagnosis Using 1D- Convolutional Neural Network Model is confusing and incorrect.

1. Line 70 - please redesign your sentence to be more clear.
2. Line 110 - please redesign your sentence to be more clear (I believe the word 'sophisticated' is not in appropriate form).
3. Line 137 - You provide the fact that Tanh is used as activation function, but no clarify the reason of such a decision.
This is mentioned later in the paper, but in my opinion should be also noted here in short term.
4. Did you use normalization (or batch-normalization) technique in some manner?
5. Line 188 - How did you choose the input vector size to be 1024?
6. How did you choose hyperparameters of the 1D-CNN?
7. You call your method an Improved 1D-CNN. I'm not sure if this is correct, since you used 'out of the box' 1D-CNN and adjusted hyperparameters. If you call something 'Improved',
it should be clear to the reader as to what and in what way. I hope that you will be able to clarify that.
8. Line 230 - The sampling frequencies should be corrected (12kHz, 48 kHz).
9. Line 258 - Relu is not suitable for training with a small number of samples. Is this literature fact?
10. Line 298 - Please clarify 0123HP term (firstly mentioned in line 283).
11. Chapter 4.3 - If you use 'Improved 1D-CNN', can it be compared to 'Non-improved 1D-CNN' and to 2D-CNN in terms of results?
12. Line 380-387 - t-SNE - you stated that "the improved 1D-CNN visualizes the characteristics". Well does it? What did you represent in Figure 15 exactly? This should be clear.
It can be seen that many authors are using t-SNE since it is great for visualization. However, t-SNE representation has to be explained correctly.
13. There are some other relevant publication that you did not mentioned, e.g.:
a) Shenfield, A.; Howarth, M. A Novel Deep Learning Model for the Detection and Identification of Rolling Element-Bearing Faults. Sensors 2020, 20, 5112.
b) Kolar, D.; Lisjak, D.; Pająk, M.; Pavković, D. Fault Diagnosis of Rotary Machines Using Deep Convolutional Neural Network with Wide Three Axis Vibration Signal Input. Sensors 2020, 20, 4017.

Author Response

Thanks for the reviewer's suggests.

Each comment has been addressed carefully in our revised manuscript and the modifications are highlighted in red.

Author Response File: Author Response.pdf

Reviewer 3 Report

The authors proposed a comprehensive model for the wrong diagnosis using a single-dimensional CNN model. The work seems interesting and achieved promising results. Before, it is ready for acceptance authors to need to address the following issues

1. Please check the title? In my opinion, “An improvement in fault diagnosis-------could be better
2. In the introduction, line 54-55 need a citation. I would like to suggest the following article

“Battineni, G.; Sagaro, G.G.; Chinatalapudi, N.; Amenta, F. Applications of Machine Learning Predictive Models in the Chronic Disease Diagnosis. J. Pers. Med. 2020, 10, 21”.    

3. Line 73-74 (citation missing), “The CNN yields satisfactory results in pattern recognition and 74 extracts features from signals or images” how this conclusion has drawn? Is there any established research work?
4. Methods should be mentioned in the documents. May I assume that section 3 can comes under the method used to achieve this result
5. Is the dataset used in this work is openly available and how this data has been extracted? Please explain
6. The results section is presented and with a comparison of other model’s authors proved this model outperformed others
7. In some areas, grammatical corrections are needed        

Author Response

Thanks for the reviewer's suggests.

Each comment has been addressed carefully in our revised manuscript and the modifications are highlighted in red.

Author Response File: Author Response.pdf

Reviewer 4 Report

Article:  An improved of fault diagnosis using 1D convolutional neural network model.

The research work discusses the possibilities of the use of vibration measurement and convolutional neural networks in the detection process of bearings faults. The language handled by the authors is not clear. The representation of the method, experimental verification, results as well as introduction contains many mistakes. Given the case, the manuscript could not be accepted in the present format for publication. The remarks are presented below:

Remarks:

Page	Line	Remark
1	30-31	The sentence “ 45-55% of mechanical … “ is unclear.
2	39-41	The sentence “A signal-based… “ is unclear.
6	176	I think you should use “low speed” not “slow speed”.
8	Fig. 3	The activation function should be shown on the scheme.
9	230	The frequency units have to be unified.
11	Fig. 5	There is a lack of units on X, Y labels. The figures should be corrected.
12	Fig. 8	There is a lack of units on X, Y labels. The figures should be corrected.
20	Fig. 15	There is a lack of units on X, Y labels. The figures should be corrected.

 

Remarks:

• What were the sizes of training, validation and testing packages for the convolution network (the structure of the data)? Are the training data and the testing data overlapping?
• Currently, there are a lot of articles describes rolling bearing diagnosis based on CWRU data. Accordingly, what is the contribution in terms of novelty?
• During the CNN training process, the authors used the ADAM method. In practical applications, learning CNN takes place by the SGD method with momentum coefficient, in which the gradient is averaged based on random learning data packets. What was the reason for using the ADAM method?
• The authors presented the operation of the diagnostic system for the electric motor based on a small number of samples. How are learning, validation and testing data differentiated?
• The authors presented verification only for a limited group of load torque, which limits the correct verification of the proposed method.
• In line 77 Authors claim that the 1D-CNN is simpler than 2D-CNN network and has lower computational complexity. The Authors should prove that statement based on the number of computation necessary to calculate the convolution operation. Moreover, Authors should explain, why the calculation 2D matrix (with the same number of elements) is slower than for 1D matrices.
• In my view the using of 4-pages theoretical background is unnecessary. Moreover, some parts are repeated in articles. The authors should correct this chapter.
• The sigmoidal function used in the structure is not efficient, as was proved by many researchers. Instead, we use ReLU nonlinearity in connection with a convolutional neural network.
• The proposed structure of CNN is very extended in comparison with the structures presented in similar articles. Authors should prove/described the process of selection of the NN structure.
• In the article, there is no information about the used programming environment, the method of implemented CNN in the diagnostic application, motor parameters.
• How the elements of batch-packets were chosen? Authors present the influence of batch size on the system accuracy based only on one probe. To show the reliable dependence a lot of different tests should be carried out.
• Presented on figure 5 the vibration signal amplitudes for ball fault are unclear. Why the amplitudes of signal for 0.021inch are significantly less than for 0.07 inch as well as 0.014 inches?
• Presented on figure 5 the vibration signal amplitudes for ball fault are unclear. Why the amplitudes of signal for 0.021inch are significantly less than for 0.07 inch as well as 0.014 inches?
• There is no quality index to evaluate the effectiveness of detection and classification.
• Why the batch size is not the power of two (2ⁿ)? This is one of the principles methods to speed up the training process.
• There is a lack of information about the hardware used in the studies (implementation of the CNN).

 

Comments for author File: Comments.pdf

Author Response

Thanks for the reviewer's suggests.

Each comment has been addressed carefully in our revised manuscript and the modifications are highlighted in red.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

The paper describes a method for faults classification in roller bearings using Convolutional Neural Networks (CNN). The method is tested using the CWRU dataset and the architecture of the CNN is adapted for handling a 1-D input signal.

The paper still has the problem of missing the cross-validation stage. Authors do not include any discussion concerning this subject. Additionally, the machine learning models used for comparison did not include any careful procedure for selecting their parameters. In consequence, authors cannot claim that their method provides better results because the comparison is unfair.

A list of issues in the paper is presented below:

1) The redaction of the following paragraphs is not clear and must be improved (lines 37-39):

“Signal - based methods from time - domain signal [5], frequency - domain signal 37 [6] and time - frequency domain [7], to the use of the signal analysis technology, the feature 38 information is extracted and fault diagnosis is made.”

Similarly:

Lines: 42-48

2) Lines 62-75 and 85-104. Authors discusses several application of deep learning techniques for fault detection. They briefly describe what the referenced authors did in the research, however, the outcome of the cited research should be describe. Was the research successful? What was the accuracy attained?

3) Before eq. (3) a reference should be given where the flatten layer and equation was described and discussed. Similarly, for the softmax function in eq. (4).

4) The phrase in lines 198-199 is not clear: “The 1D-CNN network model in this paper is obtained with reference to the Settings of other network models in the same type of research papers and several relevant tests”

If the model has been reported elsewhere, then authors must cite that research.

5) The subtitle of section 3.2 should be modified.

6) The parameters for the machine learning algorithms used for comparison seems to be set arbitrarily and there is not a systematic selection using the same method used for selecting the parameters for the proposed algorithm. In consequence, the comparison is not fair. Authors should systematically select the parameters using the same methodology for all the algorithms compared. Why the adam method is used?

7) All the machine learning methods compared should be cross-validated using the same methodology.

8) In section 3.2, authors present the description of the experiment performed but they also present in this section the results. Authors should consider to present the results in a separate section.

9) Authors do not mention any pre-processing of the data, however some of the methods compared improve their performance using normalization. Authors should verify the utility of normalization in the different methods compared.

10) The paper needs improvements concerning the readability. Some phrases are very long and difficult to understand.

 

11) The following details must be corrected

Line 266: “to to”- to

Line 271: “ithe” - the

Line 307: “Containment of 10 bearing fault types under 307 four loads of 0HP, 1HP, 2HP and 3HP)”, the phrase must be corrected.

 

 

 

 

Author Response

Thanks for the reviewer's suggests.

Each comment has been addressed carefully in our revised manuscript and the modifications are highlighted in blue.

Author Response File: Author Response.pdf

Reviewer 3 Report

Thank you for addressing our comments. The work is comprehensively improved and ready to be published. 

Author Response

Thanks for the reviewer's suggestions. We will carefully check our revised manuscript and the modifications are highlighted in blue.

Author Response File: Author Response.pdf

Reviewer 4 Report

Article:  Rolling Bearing Fault Diagnosis Based on 1D-Convolutional Neural Network

The article requires changes to the description of the research method and research results. Authors should refer to the following comments:

• What is the novelty of the proposed method of rolling bearing fault diagnosis?
• The authors suggest that "Our method is computationally less expensive than other rolling bearing diagnosis detection methods because we utilize 1D-CNN but our performance is not affected because we use Tanh as our activation function." The Authors explanations about the reason for choosing the Tanh function is not clear.

 

Author Response

Thanks for the reviewer's suggests.

Each comment has been addressed carefully in our revised manuscript and the modifications are highlighted in blue.

Author Response File: Author Response.pdf

Round 3

Reviewer 1 Report

The paper describes a method for faults classification in roller bearings using Convolutional Neural Networks (CNN). The method is tested using the CWRU dataset and the architecture of the CNN is adapted for handling a 1-D input signal.

The research topic is important and the indicated corrections have been performed by the authors.

The paper could be accepted in the present form.

Reviewer 4 Report

Article: Rolling Bearing Fault Diagnosis Based on 1D-Convolutional Neural Network

The mistakes in the representation of the method, experimental verification, results as well as introduction were corrected. Given the case, the manuscript could be accepted in the present format for publication.