Next Article in Journal
Design and Test of a New Type of Overrunning Clutch
Previous Article in Journal
Kinematic Optimization Design and Performance Simulation of Novel 5-DOF Parallel Machining Robots with Spatial Layout
 
 
Article
Peer-Review Record

Research on the Gearbox Fault Diagnosis Method Based on Multi-Model Feature Fusion

Machines 2022, 10(12), 1186; https://doi.org/10.3390/machines10121186
by Fengyun Xie 1,2,*, Hui Liu 1, Jiankun Dong 1, Gan Wang 1, Linglan Wang 1 and Gang Li 1
Reviewer 2:
Reviewer 3: Anonymous
Machines 2022, 10(12), 1186; https://doi.org/10.3390/machines10121186
Submission received: 8 November 2022 / Revised: 6 December 2022 / Accepted: 7 December 2022 / Published: 8 December 2022
(This article belongs to the Section Machines Testing and Maintenance)

Round 1

Reviewer 1 Report

1. In Equation 1, centre dot should be used (as to define function down) instead of full stop '.'. Similarly in the next sentence in order to define function f.

2.In Equation 3, do you mean multiplication or convolution? If multiplication then please use 'x' symbol.

3. Remove the left curly bracket in Equation 4.

4. In Figure 6, the activation function rectified linear unit Relu is not defined.

5. Overall analysis of the results can be improved. Discussion on successful and failure recognition should be made. No insight was given on the 1.7% failure in terms of what scenario etc.

6. Overall grammar can be improved, especially on the usage of 'a' and 'the'.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

The paper deals with the diagnosis of gearbox faults from vibration signals from fusion of time indicators and CNN.

 

The paper is interesting and allows highlighting the fusion of different methods (temporal and neural network)

The authors conclude with a model that has a recognition rate that is 98.3% high above the other methods presented in the article. This is interesting compared to traditional statistical methods. It would be interesting to put the computation times or the processing complexity of each algorithm in order to be able to decide on the real gain of the approach presented in the paper.

The presentation of the results is not clear, in particular, it is difficult to distinguish the results of each method. Authors should present the result before and after the fusion.

 Would it be possible to indicate the reduction ratio of the gearbox and to justify the use of an operating point at 900 rpm with no load.

In general, it would be interesting to give an argument on the choice of adjustment parameters for 1DCNN, dimension of the feature set, 1024 sample dimension, number of time-frequency features, IPSO-SVM parameter…

Is there a normalization of the dataset before entering them into the 1DCNN?

After equation (9), it is stated that the data can be of different dimensions before fusion. Would it be possible to make a table showing the different dimensions observed in the tests?

In Figure 5, is it possible to add the dataset {X2}?

Figure 7 and 8, should you indicate the units in g or indicate the accelerometer sensitivity?

Could the authors explain is the relationship between the data in table 4 and 5 and what are the real expected results?

Why do the authors not make the dataset available? in particular through the Data Availability Statement sections

 

A few remarks:

- there is a lack of space before figure 1.

- The equation number (9) appears twice

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

Check that the style of writing is in the third person throughout. Don’t use ‘we’.

The methodology adopted for signal processing is correct, but what is the novelty here? It is unclear.

The introduction and literature are not informative regarding signal processing. Some recent papers can be discussed regarding the use of different algorithms to classify dynamic data distribution. Some recent references can be cited, such as Health Monitoring of Turning Tool through Vibration Signals Processed using Convolutional Neural Network Architecture, Multi-point face milling tool condition monitoring through vibration spectrogram, and LSTM-autoencoder.

It is unclear whether the faults, such as wear, pitting, and broken teeth, were made artificially while acquiring data or if they are actual.

How to eliminate the possibility of misclassification of a normal tooth and a faulty tooth depending on the degree of fault?

How to eliminate the possibility of misclassification of wear, pitting, and broken teeth as normal teeth depending on the degree of fault (type II error)? If the model is deployed in real-time and such a situation arises, how will you identify that the tooth is in the failure zone and showcased as normal by your system?

How to deal with the data diversity of the present moment and moment in the future?

How to ensure the robustness of the model in a heavy noise environment?

Hyperparameters of the neural network must be summarised in tabular form.

 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 2 Report

No comments

Reviewer 3 Report

The authors have addressed all my comments carefully. Congratulations and all the best.

Back to TopTop