Nakagami-m Fading Channel Identification Using Adaptive Continuous Wavelet Transform and Convolutional Neural Networks
Round 1
Reviewer 1 Report
This paper presents a way to do channel identificaiton, i.e. identifying the type of channels, in the context of wireless communications. While channel identification is potentially useful, this paper offers no new insights to this problem and is mostly combining existing ideas with attentions to detals. There are several issues the authors need to address.
1. The author claim that the channel data generated by a channel emulator is more realisitc than data generated by MATLAB. Well, I beg to differ. What is the big difference between a "hardware simulation" and "software simulation"?
2. In apply NN models to image classification, we usually can see what "features" these NN models learn from the dataset and those features are often interpretable. In this case, it offers no such expalnation and I think it is disappointing. We don't even have a figure showing what the channel looks like in the T-F plane. Sometimes visualized data may give people inspirations. Maybe we can learn one thing or two from visualizing the channels and why some TFDs work better than others.
3. Are the authors sure that the phase, besides the magnitude of the TF representations, not necessary in this applications? If so, why?
4. It is often argued that deep learning works best from end to end; that means, it works best with raw data. In this case, the raw data is the output of the channel emulator. In this paper, the authors argue that no, the raw data should be proceesed into TFDs and then the CNN can work better. Can the authors elaborate on why this is the case?
English writing is good, except sometimes the authors spent too much time on describing very basic stuff.
Author Response
Dear Editor,
We are thankful to the reviewers for their comments and the time they have taken to read and evaluate the submitted manuscript. In this response letter, the authors address the comments by the reviewers and they explain how they have updated the new version of the manuscript on the basis of the comments.
The comments by the reviewer are highlighted with the keyword Reviewer) and the response by the authors are highlighted with the keyword Author).
Reviewer 1
Comments and Suggestions for Authors
Reviewer1 ) This paper presents a way to do channel identification, i.e. identifying the type of channels, in the context of wireless communications. While channel identification is potentially useful, this paper offers no new insights to this problem and is mostly combining existing ideas with attentions to details. There are several issues the authors need to address.
Authors) We are quite thankful to the reviewer for the time and effort to read the manuscript and to provide the comments.
Reviewer) 1. The author claim that the channel data generated by a channel emulator is more realistic than data generated by MATLAB. Well, I beg to differ. What is the big difference between a "hardware simulation" and "software simulation"?
Authors) Thank you for the comment. We agree that the sentences were not properly formulated and they were removed in the new version of the manuscript.
Reviewer) 2. In apply NN models to image classification, we usually can see what "features" these NN models learn from the dataset and those features are often interpretable. In this case, it offers no such explanation and I think it is disappointing. We don't even have a figure showing what the channel looks like in the T-F plane. Sometimes visualized data may give people inspirations. Maybe we can learn one thing or two from visualizing the channels and why some TFDs work better than others.
Authors) Thank you for the comment. We agree that showing the initial source data (time domain) under analysis and the related time-frequency transforms does already provide a perception on why the CNN is able to exploit the differences to classify the different fading conditions. We have added graphs in Section 3.2 with Figure 3 to show an example of the original weather radar signal and the related Continuous Wavelet transforms.
Reviewer) 3. Are the authors sure that the phase, besides the magnitude of the TF representations, not necessary in this applications? If so, why?
Authors) Thank you for the comment. Yes, we wrote this statement in the paper in an unclear way because it is referred only to the time frequency transform which are complex (e.g., STFT). Regarding the continuous wavelet transforms, in this study we used real mother wavelets and the CWT output is real as well. Then, we changed the paragraph and we added a clarification.
Reviewer) 4. It is often argued that deep learning works best from end to end; that means, it works best with raw data. In this case, the raw data is the output of the channel emulator. In this paper, the authors argue that no, the raw data should be processed into TFDs and then the CNN can work better. Can the authors elaborate on why this is the case?
Authors) Thank you for the comment. We added an explanation in the new version of the paper why the application of TFD is able to enhance the performance of the CNN in comparison to the application of CNN to the original time domain information. A possible explanation is that the TFD is able to highlight specific patterns which are exploited in a more effective way by the CNN. This explanation is supported by the provided results and we added references in wireless communication literature where the combination of TFD and CNN has provided superior classification to the combination of CNN and the original raw data in the time domain.
Reviewer) Comments on the Quality of English Language
English writing is good, except sometimes the authors spent too much time on describing very basic stuff.
Authors) Thank you for the comment. We have shortened some of the textbook descriptions of the algorithms, which have been used in this study.
Reviewer 2 Report
Well designed paper. The goal precisely defined - development of channel identification methods to be applied for weather radar signals. The paper not only introduces new analysis methods, but the Authors also provide thorough experimental verification with the use of laboratory setup. The conclusions are well supported by the results.
Minor recommendations:
56-59 The fading channel conditions have been created using the channel emulator in the radio frequency laboratory of the authors. In particular, we have reproduced 3GPP-like fading models based on different Tapped-Delay Line (TDL) settings with Nakamami-m fading.
It is recommended to briefly explain the reason for selection of specific fading models.
203 training and test portions of the data set are selected from the data with the same level of SNR
What was the split ratio between training and testing data sets?
Table 4 The formatting of column/row headers makes it difficult to read the data.
It is also recommended to consider publishing the datasets (for the purpose of research reproducibility).
Quality of language is good, some minor mistakes can be found.
34 The research activities in channel identification spans many -> span (activities ... span)
44, 65 On the other side -> On the other hand
516 SRN -> SNR
565 db -> dB
Table 4 db -> dB
590 optimazione -> optimization
Author Response
Dear Editor,
We are thankful to the reviewers for their comments and the time they have taken to read and evaluate the submitted manuscript. In this response letter, the authors address the comments by the reviewers and they explain how they have updated the new version of the manuscript on the basis of the comments.
The comments by the reviewer are highlighted with the keyword Reviewer) and the response by the authors are highlighted with the keyword Author).
Reviewer 2
Comments and Suggestions for Authors
Reviewer) Well designed paper. The goal precisely defined - development of channel identification methods to be applied for weather radar signals. The paper not only introduces new analysis methods, but the Authors also provide thorough experimental verification with the use of laboratory setup. The conclusions are well supported by the results.
Authors) We are quite thankful to the reviewer for the time to read the manuscript and the nice comments.
Minor recommendations:
Reviewer) 56-59 The fading channel conditions have been created using the channel emulator in the radio frequency laboratory of the authors. In particular, we have reproduced 3GPP-like fading models based on different Tapped-Delay Line (TDL) settings with Nakamami-m fading.
It is recommended to briefly explain the reason for selection of specific fading models.
Authors) Thank you for the comment. We have added a justification on why these fading models have been used.
Reviewer) 203 training and test portions of the data set are selected from the data with the same level of SNR
What was the split ratio between training and testing data sets?
Authors) Thank you for the comment. We agree that we did not provide information on the ratio and if we used a K-fold algorithm for the generalization of the results. This information was added in the new version of the manuscript.
Reviewer) Table 4 The formatting of column/row headers makes it difficult to read the data.
Authors) Thank you for the comment. We have changed the format of the table to make it more readable.
Reviewer) It is also recommended to consider publishing the datasets (for the purpose of research reproducibility).
Authors) Thank you for the comment. We have considering to make the data public and we started the internal administration process in our organization to make it public but the internal approval may take some time. We have added this clarification in the manuscript in the section Data Availability Statement in the last pages of the manuscript.
Reviewer) Comments on the Quality of English Language
Quality of language is good, some minor mistakes can be found.
34 The research activities in channel identification spans many -> span (activities ... span)
44, 65 On the other side -> On the other hand
516 SRN -> SNR
565 db -> dB
Table 4 db -> dB
590 optimazione -> optimization
Authors) We are thankful to the reviewer for the time to read the manuscript and identify these mistakes. We have corrected the indicated errors and we have also reviewed the entire manuscript for similar errors.
Reviewer 3 Report
In this manuscript, the authors investigate the channel identification problem (typical problem in wireless communications), establishing their analysis on machine learning (ML) methods. In specific, they use convolutional neural networks (CNNs) to process the time/frequency transform (e.g., spectrogram, wavelet transform, etc.) of the wireless signal as images. They also validate their methods with emulated weather radar pulse signals generated in the lab, utilizing several 3GPP-like fading models, illustrating the potency of CNNs for this problem, as well as the selection of the optimal mother wavelet for improved results. The research topic is interesting and well-presented. In my view, the following minor comment will further improve the quality of the manuscript:
· In my understanding, part of the authors’ approach has been adapted from previous analyses, since several papers can be found in the literature in this field, e.g., concerning channel state information (CSI) estimation based on CNNs. Therefore, the authors need to better highlight in the introduction the novelties of the presented approach.
Author Response
Dear Editor,
We are thankful to the reviewers for their comments and the time they have taken to read and evaluate the submitted manuscript. In this response letter, the authors address the comments by the reviewers and they explain how they have updated the new version of the manuscript on the basis of the comments.
The comments by the reviewer are highlighted with the keyword Reviewer) and the response by the authors are highlighted with the keyword Author).
Reviewer 3
Comments and Suggestions for Authors
Reviewer) In this manuscript, the authors investigate the channel identification problem (typical problem in wireless communications), establishing their analysis on machine learning (ML) methods. In specific, they use convolutional neural networks (CNNs) to process the time/frequency transform (e.g., spectrogram, wavelet transform, etc.) of the wireless signal as images. They also validate their methods with emulated weather radar pulse signals generated in the lab, utilizing several 3GPP-like fading models, illustrating the potency of CNNs for this problem, as well as the selection of the optimal mother wavelet for improved results. The research topic is interesting and well-presented.
Authors) We are thankful to the reviewer for the time to review the manuscript.
In my view, the following minor comment will further improve the quality of the manuscript:
Reviewer) In my understanding, part of the authors’ approach has been adapted from previous analyses, since several papers can be found in the literature in this field, e.g., concerning channel state information (CSI) estimation based on CNNs. Therefore, the authors need to better highlight in the introduction the novelties of the presented approach.
Authors) Many thanks for the comment. We agree that this part was not well described. We have added a new paragraph in the literature review section 1 where we summarize the key contributions of this paper against the existing literature.
Round 2
Reviewer 1 Report
My concerns have been addressed.
The writing is ok.
