Dear Authors (CC to Associate Editor, Reviewers),
Thank you very much for submitting manuscript No. TII-20-4207: "UAV-Based Sensor Nodes Localization Scheme utilizing Shallow Neural Network" to the IEEE Transactions on Industrial Informatics as a Regular Paper submission.
The review process of your manuscript referenced above has been completed. Much to my regret, I have to inform you that in the opinion of the reviewers and Associate Editor in charge, the submitted manuscript is not suitable for publication in the IEEE Transactions on Industrial Informatics.
For your reference, these comments of the reviewers are enclosed.
I look forward for your future contribution.
Dr. Gerhard Hancke
IEEE Transactions on Industrial Informatics
Comments to the Author
This paper studies the UAV-based sensor nodes localization using shallow neural networks. The motivation is that the existing schemes are too complicated to be implemented in sensor networks. However, the followings points are not explained and verified correctly.
1. What scenario is considered in this paper? and why UAV is used here. usually UAV is out-door scenarios, where GPS works very well. The authors mention there is cost using GPS, however, this is not properly discussed.
2. There lacks a detailed analysis on the implementation cost of this proposed scheme.
3. the comparison schemes are too old. it is hard to judge the performance only according to these results.
4. The simulation scenarios are not explained.
Comments to the Author
This paper discusses the localization with low cost with the assistance of UAV. There is some value for this work, but I would say it is quite far from a IEEE Transaction paper. Below are some of my more detailed comments.
1. The main point is that our accuracy is not too bad with low overhead especially for communication. But how can we justify the achieved accuracy is good, like 2 meters? Better get some comparison data. For the overhead, true, we save some GPS communication workload, but how about the communication overhead with the UAV. More investigation shall be conducted. Also, the cost comes from many parts, say we have to buy UAV, and how long can this UAV fly? What will be the amortized cost? We rather buy good sensors with larger battery? Those questions are likely to be raised up by the readers, and it should be great if the paper includes more discussion on those parts.
2. The writing shall be improved. for example, some description in the system architecture actually shall be the background or the related works. Quite often the sentences are not presented in a professional way, or even with quite some grammar errors and typos.
3. Please justify our setting and assumption is possible, i.e., why 15 blocks? Sometime like the proof of the closest pair problem, where we clearly know why it is 15. Without good justification, I would think the research is just a result of some random adventures.
4. For the 2-UAV paths. For the long-term view, you let them fly both horizontally and vertically. That is fine. But could you also describe a bit about the short-term part? say I am curious about sometimes they are near and sometimes far away. Any impact?
5. Figures some of them are blur.
6. Technical descriptions are often unnecessary. Like if people do not know std, then maybe he should not read a transactions paper. Also, for the algorithm, it is not the contribution of this paper. Using much space to describe those popular and no high-tech is quite wasteful.
7. Figure 9 may should unity the y-axis range. So we can see which signal is weak and which is strong. Then we get the idea of which one is closer.
8. The comparison algorithms are outdated. And seems that the improvement does not make any big impact.
9. The precision in the table shall be better. Too many unnecessary digits are kind of the noise for readers. As long as we can make our point to say this one is good that one is not, it should be sufficient.
Comments to the Author
The authors basically apply extreme machine learning for UAV-assisted localization of the sensor nodes and claim that this method improves the accuracy of localization while also reducing the computation overhead. While localization is indeed an important problem, the scientific contribution of this paper is not significant, and the manuscript has are several weak points- both scientific and semantic. The paper lacks novelty authors simply apply ELM for the case of UAV-enabled sensor localization. Neither ELM nor UAV-assisted scenario are new. In addition, the performance evaluation and analysis is limited to an area of a total of 100m by 100m with each block of 10m by 10m. This is Certainly not sufficient to prove the effectiveness of or improvement brought by the proposed scheme. Also the writing of the paper is poor which makes it difficult to read.
Comments to the Author
A UAV- and shallow neural network-based sensor nodes localization technique is proposed and its effectiveness is demonstrated in terms of location accuracy and solution efficiency, using synthetic data.
My major concerns are the following:
1. It's unclear where the proposed algorithm will be run. The paper mentions using a "computation system other than the sensor nodes" but without providing further details. Elaborate on the computation system. Where is it located? For example, in Fig. 8, where is the SLFN training implemented? How do the sensor nodes communicate the measured RSSI to the computation system? What is the latency? What information is communicated from the computation system back to the sensor nodes? What is the latency?
2. UAV trajectory shown in Fig. 2: why constrain the UAVs to move in horizontal and vertical directions? How to prevent collisions between the 2 UAVs?
3. Received RSSI at the sensor nodes: How many receive antennas are on each sensor node? How does each sensor node distinguish between the beacon signal from the UAV left and right antennae?
4. Connection states: how long is a connection state? What is the relationship between T and \m_k shown in Fig. 5? How is the value of T determined?
5. Eqn. 3: an exponent of 2 is missing in the numerator: (R_i - \mu)^2
6. What is the difference between N and n shown in Fig. 6?
7. Instead of setting the output layer to the target output in eqn. (6) and then solving for \beta in eqn. (7), why not minimize the error between the predicted output (eqn. (5)) and the target output?
8. Network Configuration and Simulation settings are NOT specified. Hence, it's difficult to put the presented results in proper context: Please state in a Table the assumed values in your simulations of the following parameters:
a) Size of the sensor field (i.e. network area)
b) Propagation channel parameters: L_0, d_0, \alpha, \sigma
c) Transmit power by each UAV antenna
d) Number of sensor nodes in the service area
e) Number of UAVs in the service area
f) Value of \m_k, maximum number of connection states for each node k
g) Value of T, number of connection states
h) Value of N, number of training samples (i.e., size of training data)
i) Size of test data
j) Why set \N_h = 20?
9) All the results are based on synthetic data, consider using actual field data.
10) Figs. 10 & 11: What is the assumed activation function?
11) In Figs. 12 and 13, why does the hardlimit activation function perform better at the training stage but poorly at the testing phase?
12) Fig. 14: what activation function was assumed?
13) Your results show that the EML exhibits the best location accuracy and best solution efficiency compared to BP and SVM. So what is the tradeoff for the EML? I.e., what are EML's limitations?
13. Editorial/Presentation quality: The paper requires a major editorial rework. The presentation quality is very low.
Comments to the Author:
(There are no comments. Please check to see if comments were included as a file attachment with this e-mail or as an attachment in your Author Center.)email@example.com, firstname.lastname@example.org