Next Article in Journal
Secure Device-to-Device Communication in IoT: Fuzzy Identity from Wireless Channel State Information for Identity-Based Encryption
Previous Article in Journal
Domain Adaptive Subterranean 3D Pedestrian Detection via Instance Transfer and Confidence Guidance
 
 
Article
Peer-Review Record

Recognition of Ethylene Plasma Spectra 1D Data Based on Deep Convolutional Neural Networks

Electronics 2024, 13(5), 983; https://doi.org/10.3390/electronics13050983
by Baoxia Li 1, Wenzhuo Chen 1, Shaohuang Bian 1, Lusi A 1, Xiaojiang Tang 1, Yang Liu 2,3, Junwei Guo 2, Dan Zhang 2, Cheng Yang 2 and Feng Huang 2,*
Reviewer 1: Anonymous
Reviewer 2:
Reviewer 3:
Reviewer 4: Anonymous
Electronics 2024, 13(5), 983; https://doi.org/10.3390/electronics13050983
Submission received: 4 February 2024 / Revised: 22 February 2024 / Accepted: 27 February 2024 / Published: 4 March 2024
(This article belongs to the Section Computer Science & Engineering)

Round 1

Reviewer 1 Report (Previous Reviewer 4)

Comments and Suggestions for Authors

1. When we talk about neural networks, we are talking about the problem of identification (spectral curves or patterns).  This paper lacks a definition of what is actually being identified. What images are present in the specta?
2. In conclusion, it does not reflect what the development was carried out for, there is no brief description of the network features that allows you to achieve a result and it does not show how you can use the results in practice.

3.The text has not been designed, which makes it difficult to read it.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report (Previous Reviewer 3)

Comments and Suggestions for Authors

 

Some improvements to the article have taken place. However, a number of my other comments from the previous review were not taken into account by the authors of the article. I propose to give answers to such problematic questions:

1.    The procedure for selecting hyperparameters of the model for some reason does not take into account the interaction of these hyperparameters as factors (see the literature on factorial experiments) affecting the characteristics of classification accuracy.

2.    It is not clear what the authors have contributed to the theory of Deep Convolutional Neural Network? Is it only the application aspect of this class of networks that is considered here?

3.    It is not clear due to which algorithmic innovations such high accuracy rates were achieved.

4.    In the paper, it is worth describing in more detail the reasons for choosing the structure and parameters of the Deep Convolutional Neural Network. Why exactly this type of model, its architecture and parameters?

5.    It is worth comparing the accuracy, computational complexity, and interpretability of the approach proposed by the authors with classical methods of accepting hypotheses based on procedures for statistical evaluation of probabilistic characteristics of the studied stochastic spectrum. In my opinion, even the assessment of the mathematical expectation of the studied spectrum is an informative feature for recognizing the corresponding states.

6.    In general, it is not clear what is the practical value of such a classification? Why classify spectra by power levels? These levels are known to the experimenter.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report (Previous Reviewer 2)

Comments and Suggestions for Authors

In the manuscript “Recognition of Complex Plasma Spectra Based on Deep Convolutional Neural Network” resubmitted by Baoxia Li, Wenzhuo Chen, Shaohuang Bian, A Lusi, Xiaojiang Tang, Yang Liu, Junwei Guo, Dan Zhang, Yang Cheng, and Feng Huang the authors demonstrate an ethylene discharge spectrum recognition model based on the deep convolutional neural network. The authors can reach an accuracy higher than 98% under the classification based on the RF discharge power.

Compared with the previous version, the authors refined all the parts of the manuscript, modified the figures, and added more details about the training model. The manuscript overall has a better quality. This manuscript can be published with the consideration of the following remarks:

(1)   In section 2, the authors introduce the data classification method which is based on different RF powers. More reasoning is expected at this specific point. For example, what is the motivation for classifying in this way? Why select RF powers between 60 to 69 W? Also, maybe I missed it, I didn’t find the discussion about “The discrete values of RF discharge power might limit the application of the model. Could the authors comment on that?”

(2) In section 4.3, the authors introduced an f parameter to indicate the performance of each model. What is the definition of the f parameter?

(3) I still recommend considering expanding the discussion on the limitations, uncertainties, and future directions. Several more sentences in the conclusion part will very likely inspire people on the next classification method. This will help people better understand the scope of this method too.

(4) Please organize the manuscript in the uncomment view. Please double-check the whole manuscript and pay attention to any potential typos. For example, in line 33: “And it is an indispensable means of developing …”.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 4 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

The article requires revision, and I would like to highlight certain aspects for the authors' attention:

01. The procedure for selecting hyperparameters of the model does not seem to consider the interaction of these hyperparameters as factors influencing classification accuracy.

02. It is unclear what specific contributions the authors have made to the theory of Deep Convolutional Neural Networks. Is the focus solely on the application aspect of this network class?

03. The rationale behind choosing a deep convolutional neural network for ethylene plasma spectra recognition over traditional models needs more clarification.

04. The paper lacks clear explanations for achieving high accuracy rates through algorithmic innovations.

05. It would be beneficial to elaborate on the reasons for selecting the structure and parameters of the Deep Convolutional Neural Network. Why precisely this model, its architecture, and parameters?

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Round 2

Reviewer 4 Report (Previous Reviewer 1)

Comments and Suggestions for Authors

Can be accepted.

This manuscript is a resubmission of an earlier submission. The following is a list of the peer review reports and author responses from that submission.


Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have presented an article proposing a model for ethylene discharge spectrum recognition based on deep convolutional neural networks.

1. The problem statement addressed lacks novelty, with references to older research articles for problem formulation.

2. The data collection process and exploratory data analysis methods are not elucidated. A description of these aspects would enhance the comprehensibility of the study.

3. The rationale behind choosing a deep convolutional neural network for complex plasma spectrum recognition over traditional models is not sufficiently explained.

 

The readability and novelty of the paper are deemed insufficient, and, therefore, I cannot recommend it for further review.

Comments on the Quality of English Language

Minor editing of English language required.

Reviewer 2 Report

Comments and Suggestions for Authors

In the manuscript “Recognition of Complex Plasma Spectra Based on Deep Convolutional Neural Network” submitted by Baoxia Li, Wenzhuo Chen, Shaohuang Bian, A Lusi, Xiaojiang Tang, Yang Liu, Junwei Guo, Dan Zhang, Yang Cheng, and Feng Huang, the authors demonstrate an ethylene discharge spectrum recognition model based on the deep convolutional neural network. The authors can reach an accuracy higher than 98% under the classification based on the RF discharge power.

The authors first introduce the data collection and classification methods of the ethylene spectra curves. A neural network structure is constructed with the emphasis on the residual shrinkage block which aims at eliminating the redundant features. After choosing the proper learning rate, batch size, epoch, and optimizer, the authors can classify the spectra with over 98% accuracy. The proposed model is also compared with the other existing models.

Overall, the manuscript is well-written with clear structures. The results are presented and explained properly. Generalization of the conclusions could make this work more impactful. Some points need a more careful explanation. This manuscript can be published with the consideration of the following remarks:

(1)   In section 2: data collection, the authors introduce 10 classes of spectra curves. It is classified based on the RF discharge power. Is this model only suitable for this classification criteria or it is suitable for the other criterion, different ethylene pressure, for example? What is the motivation for classifying in this way? The discrete values of RF discharge power might limit the application of the model. Could the authors comment on that? Some explanations will help readers understand the logic behind it.

(2)   In section 4.3, the authors show the experimental results. However, some more information is needed to understand the model. How large is the training and testing set? What is the time scale of training one model and what is the speed of converging?

(3)   Some figures can be improved. In Figure 5, the labels of both axes are missing. In Figure 6, the labels are a little small to read. In Figure 7, the y-axis label is missing.

(4)   On page 10, the authors compare the evaluation indicators by different models. What are the running times of all those models? How is it compared with the proposed model in the manuscript? Among those algorithms, are they all trained with the same data sets? Please include the information in the manuscript.

(5)   In Table 9, the authors present the results of the proposed model applied to the public dataset. What are the criteria for classifications in these datasets? Surprisingly, the humidity dataset predicts 100% accuracy. Can the authors comment on that?

(6)   Please also consider expanding the discussion on the limitations, uncertainties, and future directions of this neural network model. This will help people better understand the scope of this method.

Reviewer 3 Report

Comments and Suggestions for Authors

The article needs revision. I would like to draw the attention of the authors of the article to the following aspects:

1.    The procedure for selecting hyperparameters of the model for some reason does not take into account the interaction of these hyperparameters as factors (see the literature on factorial experiments) affecting the characteristics of classification accuracy.

2.    It is not clear what the authors have contributed to the theory of Deep Convolutional Neural Network? Is it only the application aspect of this class of networks that is considered here?

3.    It is not clear due to which algorithmic innovations such high accuracy rates were achieved.

4.    In the paper, it is worth describing in more detail the reasons for choosing the structure and parameters of the Deep Convolutional Neural Network. Why exactly this type of model, its architecture and parameters?

5.    It is worth comparing the accuracy, computational complexity, and interpretability of the approach proposed by the authors with classical methods of accepting hypotheses based on procedures for statistical evaluation of probabilistic characteristics of the studied stochastic spectrum. In my opinion, even the assessment of the mathematical expectation of the studied spectrum is an informative feature for recognizing the corresponding states.

6.    In general, it is not clear what is the practical value of such a classification? Why classify spectra by power levels? These levels are known to the experimenter.

7.    I do not see the point in citing Table 7. It does not contribute anything informative to the research.

Reviewer 4 Report

Comments and Suggestions for Authors

1. The main drawback of the paper is the lack of justification for the neural network model. In my opinion, the neural network model looks strange because  the process begins with the convolution layer. This is the processing of local features, which themselves are filters against local noise. This technology should work well for images. But for the spectrum it causes loss of bins. I did not see the rationale for this choice of model.

 

2.  The spectrum diagnosis of complex plasmas should include noise model. Because plasma has such specification and the authors talk about pattern in spectra

 

3. Tests must be performed on special data sets that include special spectra and noise. This is not mentioned in the paper. A random set of spectra is unacceptable for solving this problem

 

4 Comparation with 246 models 1DCNN, VGG13, ResNet18 and AlexNet is also is strange.  VGG13, ResNet18 and AlexNet is old models oriented for images but they no oriented for spectrum. 

Back to TopTop