A Joint Denoising Learning Model for Weight Update Space–Time Diversity Method^†

Round 1

Reviewer 1 Report

The authors have addressed the main concerns from the review.

Reviewer 2 Report

The authors have performed significant revision and improvement of the paper compared with previous version. Therefore, the paper could be accepted in the present form.

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

Actually, I don't know the key point of this manuscript. Although author summarize their main points in L111-132 of the introduction, the description seem not to exhibit their novelty explicitly. 

 

<Major comments>

1) For the noise reduction, many deep learning algorithm based on image processing can be applied and have been presented. Why did the authors use CGAN? Does it have an advantage for the underwater environment? I don't find any explanation. 

2) In CGAN, What is the dimension of input and output data? Add it. What does 'R_yz/I_yz' mean? Is it element division? If so, as the element of I_yz -> 0, what happens?

3) Precisely describe the structure of TDNN. In his manuscript, the description of the network structure (layer dimension, input/output dimension, filter information..) is neglected or omitted. I think that it is difficult to reproduce this study.

4) In Sec 3.3, it is hard to notice the intention of the authors due to mainly the unclear English writing (Grammar or sentence is good). What is the novelty in the STD model with weight update? Is it just a kind of adaptive processing?

5) In simulation, the procedure on training the learning model has to be described. How is the noise generated? What is the definition of SNR used ? (sensor SNR? array SNR, broadband SNR?...), The size of data? the ratio of training, validation, and test data? How is the Hyperparameter  determined in the network? Almost everything are missed. 

As judged by the results and conclusion, this manuscript clearly includes a meaningful results. By the figures presented by the authors, the proposed method reduce the MSE and the BER. However, there is few description of that reason and method.

 

<minor comments>
1) In sec 2, I find three kinds of notation, y(t), bold{y(T)}, y_i(t) in equation. What is difference? And, In Eq. (1), what is x_i(t)? there is no explanation.

2) (L160 and Eq. (2) ) Eq. (2) is understood to represent s(t). It is not related to Eq. (1).

3) (L185) What does 'non-smooth random process' mean?

4) (L191-193) The sentence is not clear.

5) (Fig. 4) What does 'the second gray arrow' means? The first gray arrow seems to represent the f-k transform.

6) (Eq. (4) and (5), L214-217) The explanaton is incoplete. R_sz, I_sz, R_yz, I_yz??

7)  (Fig. 5) The caption may be wrong. It is to plot noise-reduction, restoration, and decompression processing.

8) In  Eq. (9) and (12),  there are typos.

9) (L281) there is duplicate expression.

10) (L359-361) How is the noise simulated? You considered the statistical property between sensors?

11) (Sec. 4.2) In lake environment, How is SNR?

 

Reviewer 2 Report

The paper clearly gives the environment, assumptions, requirements and objectives of the problem in hand, and points out major issues or difficulties when dealing with the problem. The paper is well-organized, but it can be improved as follows:   1. A vest amount of existing studies have proposed the similar concept to improve communication quality with machine-learning based equalizers. This paper should review the class of these papers and clearly give the differences and major improvements of the proposed mechanism.    2. The simulation part is weak and the quality of Section 4 can be improved. The discussion and evaluation part (Section 4) only considers several different parameter settings for the authors’ solution and compares this work to the conventional schemes (e.g. Space-Time Diversity schemes). Therefore, it is unclear if and how this work advances the state of the art.   3. The authors should provide a discussion on implementation complexity  (e.g. communication complexity, time complexity, and computational complexity) and performance trade-offs in the proposed denoising learning model.

Reviewer 3 Report

In this paper, it is proposed a joint denoising learning model with a weight updating Space-Time Diversity (STD) structure to improve communication quality. With this aim, a deep learning-based noise-reduction model and a STD structure applicable to underwater communication are constructed. To reduce the effect of Vessel’s Underwater Radiated Noise (VURN) on the received signal, it is proposed a noise-reduction model trained using the real and imaginary parts separately as features. To improve communication quality, the STD method, based on the weight update strategy, is proposed, which focuses on communication in the main direction to enhance communication quality.

Some shortcoming and missing of the paper are the following:

1. The following phrase is non-understandable: “This is way, adding equalization processing after STD combining is unnecessary, significantly reducing the computing process. in equation (1) can be further expressed as” (Lines 158-160).
2. There is a discrepancy of open and close brackets in formula (10); c_j^r must be replaced by v_j^r.
3. What does mean “mathbf” in formula (12)? 
4. What does the multiplier j mean in formula (13)? 
5. What does mean superscript H in formula (16) and subsequent formulae?
6. What do mean b and p in Line 302?
7. What does mean σ_p in formula (20)?
8. Theoretical part is vey long and weakly connected with the subsequent calculations.
9. Are all the formulaе in the theoretical part original? If not, then the appropriate references should be provided.
10. Conclusions should be extended for account of the results obtained.