A Multi-Modal Modulation Recognition Method with SNR Segmentation Based on Time Domain Signals and Constellation Diagrams
, Xinze Li 1,2, Haiyan Zhang 1,2,*, Guoting Yang 1,2, Shurui Li 3, Peng Cheng 4,5 and Yonghui Li 5Round 1
Reviewer 1 Report
The main achievement presented in the paper is the novel structure of the neural network for Automatic Modulation Recognition for wireless networks. Authors called their neural network M-LSCANet, which includes SNR segmentation and outliers modification. The scientific experiments are well planned and performed. All the results are well presented and explained in the paper. The overall performance of the M-LSCANet in terms of Automatic Modulation Recognition success seems to surpass the other current state-of-the-art solutions, while TMRN-GLU is the closest in the high SNR range when the net complexity is taken also into consideration. In my opinion the paper is worth to be published in the Electronics Journal. I have a minor comments concerning the editing.
They are as follows:
1. Beginning of the line 40: “Space” is missing.
2. Reference [39] is called in line 40, but we have author mismatch with the References segment.
3. References segment contains such signs as [J] or [C]\\
4. In equation (1) convolution sign should be used between s(t) and c(t).
5. Line 227: Comma should be used after “period”.
6. We have notation mismatch between equation (4) and the line below in Rayleigh fading factor.
7. In Algorithm 1, the signs used for operations between CBAM with M and W with ReLu are not explained.
8. Line 405: I think, it should be 4:1 instead of 8: 2.
9. Some titles of the text segments are not well positioned, e.g. line 408.
10. Line 443: “proposed” instead of “propose”.
11. Line 599: “Table 5” instead of “table 5”.
12: Line 606: “Space” is missing after “2.2%”.
13: Line 607: Comma is missing before “respectively”.
14: Line 642: “multi-modal” instead of “multi-model”.
Comments for author File: 
Comments.pdf
No major Issues detected.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
In this paper, the authors proposed a modulation recognition method based on machine learning techniques. The study is impressive and well matches the trend of the wireless research area. However, I have some concerns as following.
1) In this study, the authors adopt 10 modulations to test the proposed method. However, some modulations show very different constellations under different SNR. For example, in Fig. 10, the 3rd, 6th, and 10th modulations show totally different forms of constellations among (a) SNR = 12 dB and other SNRs. It is better to explain this for better understanding.
2) In Fig. 16, it is better to explain why removing outliers seems no help for the classification accuracy.
3) In Fig. 18, it is better to explain why the classification accuracy of WBFM cannot be better than 70%.
None.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
The manuscript gives an overview of automatic modulation recognition (AMR) and neural networks and machine learning for this application area. Overall, the manuscript presents an extremely complex and intellectually challenging subject that requires an in-depth understanding of the subject matter. The author has made noticeable efforts in the scientific implementation of this complex research topic and has provided valuable insights. The manuscript has a clear structure
However, there are several areas that need significant improvement to improve the clarity and comprehensibility of the manuscript for non-experts.
The manuscript can only be understood by persons with deep knowledge of the subject. The reader should have a profound understanding of the various types of neural networks and of digital and analogue modulation techniques. Most terms are only mentioned without further explanation, but some terms are explained in detail by using external literature references. But, for example, the cited literature reference, number 4, is provided in Chinese language and writing, which is not very helpful when studying.
For the understanding of the manuscript from the point of view of a non-expert, the additional study of the referenced external literature is absolutely necessary. This is a pity!
On the other hand, the proposed structural architecture of the network model is explained very extensively and in detail, but brief explanations also being missing here. Example:
Line 307:
What do the parameters in Conv2d(32,8,1,1,0) in Table 1 mean? A brief explanation would make it understandable.
Line 85: What is ResNet?
Add a literature reference or an explanation for ResNet.
Line 198: Confusing sentence:
“ ... over the entire SNR range of -20dB~18dB.”
What does “-20dB~18dB” means? Do you mean “SNR range of -20dB to +18dB.”
Line 200: Confusing sentence:
Remark: Exchange the sentence “The reminder in this paper is organized as follows” with the simple readable sentence “This paper is organized as follows”
Line 261: Literature reference in Chinese language [49]
A Chinese Literature reference is not very helpful for an English manuscript.
Recommendation: Remove all non-English literature references, as MDPI manuscripts are mainly consumed by English speaking readers.
Line 235:
“ .... where a(t) represents the Rayleigh fading factor”
Do you mean α(t) instead of a(t)?
Line 399:
Write “Prof. O’Shen” instead of “Prof. O’shen”.
Line 407:
What is WBFM? Do you mean Wide Band Frequency Modulation? Add the full version of the abbreviation in the text.
Line 407:
What is AM-DSB? Do you mean Amplitude Modulation (AM) - Double Sideband (DSB)?
Add the full version of the abbreviation in the text.
Recommendation: To improve the readability, add a legend with full version of the abbreviations, used the first time in Table 3. Like as: BPSK, QPSK, 8PSK, QAM16, QAM64, GFSK, CPFSK, PAM4, WBFM and AM-DSB.
Line 433: Confusing sentence:
“The default argument Adam is used as the optimizer for learning hyper-parameters.”
What is Adam?
After internet research I found a usefully document, which explain the Adam optimizer algorithm.
Please reformulate the sentence and provide a literature reference, like as:
D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” Available: http://arxiv.org/abs/1412.6980
Line 569: Confusing sentence:
“This can be attributed that AM and FM signals in the dataset at the same carrier frequency are in the silent period [24].” What is silent period?
After brief research in the referenced paper [24] it is more clearly explained what is meant:
Excerpt from the referenced literetaure [24]: “In the case of WBFM/AM-DSB the analog voice signal has periods of silence where only a carrier tone is present making these examples indiscernible.”
On line 590 you provide a clearer explanation.
Remark: reformulate the sentence and provide the sentence from line 590 at line 359.
Recommendation: The research results are presented extensively and in detail. – But Figure 8, Figure 10, Figure 12 and Figure 13 should be presented using a table with the modulation modes (8PSK, 428 AM-DSB, BPSK, CPFSK, GFSK, PAM4, QAM16, QAM64, QPSK and WBFM) on top of the table and the constellation diagrams underneath.
Not all used abbreviations are supported in the long version, which severely limits good readability.
After my evaluation, the manuscript is written in a well readable and understandable English
As already mentioned, the implementation work is presented in great detail, but nevertheless, it is an enormous advantage to have prior knowledge of automatic modulation recognition (AMR) and a deep knowledge in neural networks and machine learning technologies. It would be very helpful if the components of the neural network model were described in a bit more detail in the introductory chapter.
From my point of view, the manuscript can only be published after a major revision.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 4 Report
I think the work is good however I have the following comments:
1. Why is -4dB set as the threshold? Is this based on energy detection? what is the theoretical reason behind it?
2. Why did the authors use time-domain signals as the input? This might be good in noise-limited scenarios such as the work in:
10.1109/VTC2021-Spring51267.2021.9448711
or
10.1109/LCOMM.2022.3217337
however this is disputed for interference scenarios such as the work in:
10.1109/TMLCN.2023.3270131
so can the authors elaborate on their choice?
3. Again, what happens to this methodology under high-interference settings?
English is good
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 3 Report
After the major revision of the manuscript, the manuscript represents a readable and understandable work even for a non-expert. All comments and recommendations have been considered and from my point of view the manuscript can be released through this significant revision.
