kumoh national institute of technology
Networked Systems Lab.

Rubina "RFDOA-Net: DOA Estimation of RF Signal for UAV Localization Based on Residual CNN"
By : Rubina
Date : 2020-06-29
Views : 155

From:

wlzhang@mail.xjtu.edu.cn

To:

EIC.COMML@l2s.centralesupelec.fr, adm.comml@l2s.centralesupelec.fr, wlzhang@mail.xjtu.edu.cn

CC:

Subject:

Comm Letters CL2020-1110 Rejected - No resubmission allowed

Body:

Dear Author(s):

The review of the referenced manuscript, CL2020-1110, entitled RFDOA-Net: DOA Estimation of RF Signal for UAV Localization Based on Residual CNN, is now complete. I regret to inform you that based on the enclosed reviews and my own reading of your manuscript, I am unable to recommend its publication in IEEE Communications Letters.

Your paper may not be resubmitted for review. The reasons for this are as follows: The reviewers questioned the novelty and contribution of this paper. The authors just establish a deep with more layers than that in existing methods neural network to deal with the DOA estimation problem. The comparison between the proposed method and the classic model driven methods has not been provided. Multipath effect is another important issue that should be taken into consideration.


Additional comments include: It is not clear how many waves have been considered in the numerical results. Moreover, it is not clear whether the proposed network rely on the prior knowledge of P. If required, what about the robustness to the imperfect knowledge of P.



The reviewers' comments are found at the end of this email.

Thank you for submitting your work to the IEEE Communications Letters.

Regards,

Dr. Weile Zhang
Associate Editor
IEEE Communications Letters

Reviewer: 1

Comments to the Author
See Attached File

Reviewer: 2

Comments to the Author
In the past few years, many researchers in the area of DOA estimation are turning their attention to deep learning techniques. DL indeep provides a possible way for addressing some particular DOA estimation problems. Readers will appreciate it if papers can provide some insights to show what the particular problems are and why DL techniques are useful in addressing them.
However, this manuscript does not provide such contributions. The authors just establish a deep (with more layers than that in existing methods) neural network to deal with the DOA estimation problem, and then demonstrate its effectiveness via some specially designed simulations. A great amount of doubts thus arise, e.g., why UAV? why multipath? why residual CNN? why so many layers? how will the framework behave if the scenario changes, including multipath/DOA/array parameters? and etc.
From the perspective of this reviewer, this manuscript does not provide enough useful information for researchers in this area.

Reviewer: 3

Comments to the Author

1. How the proposed approach differs from existing methods, and how this impacts performance.

2.please describe how the power attenuation and the time delay are set during the simulation


3.The superiority of the proposed method over the classic model driven e.g. MUSICmethod has not been clarified.

4.How to deal with the problem of angle ambiguity when the array is NULA?

5. Source code is not available. I believe it should be published for a deep evaluation of the work since the experiment data is synthetic. If there is some reason for not make it public, it should be described in the paper.


Reviewer: 4

Comments to the Author
In this manuscript, the authors utilize an elaborately designed neural network to estimate the DOA of the received RF signals. There are some issues to be addressed before it would be published. I can provide comments as follows.

1. The authors provide three major contributions in the introduction. However, the second and the third points are about the experimental results, which might not be regarded as major contributions.

2. In this manuscript, the work of estimating DOA is formulated as a classification task rather than a regression task. It is suggested to explain whether the estimation accuracy is acceptable when considering the quantization error. Besides, the range of the angle is set as -80 degree to 80 degree, which might not be practical.

3. The authors claim that their method considers the channel impairment introduced by multipath propagation. However, the evaluation results have not covered the effects of multi-paths. For example, how do the authors set the number of paths in their experiment.

4. Also about the effects of multi-paths. Since there are more than one paths when considering the NLOS, each path will contain a DOA. Therefore, how do the authors ensure the estimation accuracy of the LOSs DOA if the received signal is biased. The mechanism should be elaborated in detail.

5. The writing is suggested to be further polished. Some sentences are too long.


Reviewer: 5

Comments to the Author
This letters proposes an DOA Estimation method based on residual CNN and provides its performance in terms of estimation accuracy and RMSE. Generally speaking, the residual CNN structure is a mature technique that has been applied in many existing works, thus, this review feels the novelty and contribution of this letter are limited. It is not clear why this technique must be used for the DOA estimation problem. Is it necessary to use residual CNN for this DoA estimation problem? What is the performance if we simply remove the residual branch in the current network structure in Fig. 2?

Other comments:

1. This letters only provides some performance results with different hyper-parameters settings, but lacks of the explanation, analysis, or insights how to set these hyper-parameters, such as filter size, number of filters, input dimensions. Table II actually does not contain too much useful information, but only shows that the performance is better when there are more trainable parameters, which is common in all works related to neutral networks.

2. The two existing methods for comparison need some data preprocessing, but the proposed method does not. It is not clear what the performance improvement comes from. Is it due to the residual network structure or due to the direct processing to the raw data without any preprocessing?

3. It is not clear what Eq.(4) means and what the connection between it and Eq.(2).

4. It seems not fair to compare NULA with ULA with two different apertures. Better to compare them with the same aperture.

5. It would be better to open some data and source code, which would greatly increase the reliability of the simulation results.

6. In the simulation, it seems neither the accuracy nor RMSE is defined. The setting of multipath propagation is not explained either.
Attachments
Attachment 1:   ÷ review.pdf(26.6KB)