Deep Learning-Based End-to-End Carrier Signal Detection in Broadband Power Spectrum
Round 1
Reviewer 1 Report
In this paper, an end-to-end deep convolutional neural network model is proposed for carrier signal detection in the broadband power spectrum. The model is meaningful, however, there are some issues should be addressed in current version.
1. The novelty of the proposed model should be highlighted in Abstract.
2. The relationships between Section 2, 3 and 4 are not clear.
3. The conclusions obtained from Fig.6 and 7 should be clearly presented.
4. Based on Table 1, it can be found that compared with FCN method, the SCN method improves the AP from 98.32% to 99.88%, AR from 98.13% to 99.12%, F-Score from 98.22% to 99.48%, respectively, and the superiority is not clear. In addition, the proposed method is time consuming than the FCN method.
5. More compassion results with other methods should be presented.
Author Response
Response to Reviewer 1 Comments
In this paper, an end-to-end deep convolutional neural network model is proposed for carrier signal detection in the broadband power spectrum. The model is meaningful, however, there are some issues should be addressed in current version.
Point 1: The novelty of the proposed model should be highlighted in Abstract.
Response 1: Thank you very much for the comment. We updated the Abstract to highlight the novelty of the proposed method in the manuscript. We presented it as below:
Abstract: This paper provides an end-to-end deep convolutional neural network (CNN) model for carrier signal detection in the broadband power spectrum, named Spectrum-Center-Net (SCN). Regarding the broadband power spectrum as a one-dimensional (1D) image and the sub-carriers are target objects on the image, the carrier signal detection problem's core task turns into the frequency centres (FC) and bandwidths (BW) regression. Our designed SCN takes the broadband power spectrum as inputs and extracts the features of different length scales by the ResNets backbone. Then feature pyramid network (FPN) neck fuses the features and outputs the fusion features. Next, the RegNet head regresses the power spectrum distribution (PSD) predictions for FC and the corresponding BW prediction. Lastly, we can get the sub-carriers targets by applying non-maximum suppressions (NMS). We train the SCN on a simulation dataset and evaluate it on an actual satellite broadband power spectrum set. As an improvement of the fully-convolutional-network-based (FCN-based) method, the proposed method directly outputs the detection results without post-processing. Extensive experimental results demonstrate that the proposed method effectively handles the carrier signal detection in the broadband power spectrum and achieves higher and more robust performance than the deep FCN-based methods and threshold-based methods.
Point 2: The relationships between Section 2, 3 and 4 are not clear.
Response 2: It is very grateful that you pointed out that the relationships between Section 2, 3 and 4 are not clear. We updated the structure of the manuscript, combined Sections 3 and 4 into one section: Methodology which includes 3 parts, "SCN Architecture", "SCN Training Targets and Loss Function" and "SCN Inference Details"; and modified some titles of the original Sections 2, 5 and 6, to make the article structure more transparent. And now the titles of each section are presented as below:
- Introduction
- Problem Description
2.1. The Core Task of Carrier Signal Detection Problem
2.2. The End-to-end Detection processes
- Methodology
3.1. SCN Architecture
3.2. SCN Training Targets and Loss Function
3.3. SCN Inference Details
- Experiments
4.1 Data Preparation
4.2 Model Training
4.3 Evaluation Results
4.4 Performance Comparison to other methods.
- Discussion and Conclusion
Point 3: The conclusions obtained from Fig.6 and 7 should be clearly presented.
Response 3: We greatly appreciate your comment. We updated the conclusion section and added some discussion of the results of some of Figures and Table, including Fig. 6 and 7. Details are as follws:
In summary, we proposed an end-to-end deep learning-based method for carrier signal detection in the broadband power spectrum, called SCN. We regard the carrier sig-nal problem as an object localization task in the 1D broadband power spectrum image in-stead of the segmentation task in the FCN-based method, and the core task is to regress the centres of all sub-carriers and their corresponding bandwidths. To improve the FCN-based method, we apply the attention mechanism and use more complex residual layers in the feature extraction modules. Then the FPN neck fused the different scales of features and automated regress to PSD and BW predictions by the RegNet head. As Table 2 shows, experiments have suggested that the SCN method, by training with the simula-tion samples, effectively handles the carrier signal detection problems and achieves higher performance than the FCN-based methods and threshold-based methods.
Moreover, in the training process, Figures 6 and 7 suggest that the proposed model converges with the training epochs. Meanwhile, increasing the scale of the model or using more numbers of training samples both improve the PSD and BW loss lower. However, while the SCN scale is beyond 8, the distance of performance on the validation simulation samples gets smaller and smaller with the training epochs increase, as Figure 8 shows. Also, Figure 9 shows that the number of training samples slightly impacts the validation performance when the losses converge. Considering that the validation samples are also simulated in the same conditions as the training samples, we analyze that the model tends to overfit as the losses converge. Increasing the variety of sample generation condi-tions or the actual signal sample numbers would help alleviate this phenomenon.
Last, in the evaluation phase, Table 2 shows that the performance increases with the SCN scale. While Table 3 indicates that the SCN method needs too much computation com-plexity, which causes more inference time than the FCN-based methods, even though SCN methods do not need post-processing. So in our future work, we would try to reduce the computation complexity of the proposed method to make it more efficient in real applica-tions.
Point 4: Based on Table 1, it can be found that compared with FCN method, the SCN method improves the AP from 98.32% to 99.88%, AR from 98.13% to 99.12%, F-Score from 98.22% to 99.48%, respectively, and the superiority is not clear. In addition, the proposed method is time consuming than the FCN method.
Response 4: We appreciate your helpful comment.
First, SCN is an improvement for FCN method, mainly focusing on two points of situations which have been mentioned in the introduction Section (P2). And we revised the manuscript in the last of Section 4.4 (P12) to highlight the improvements. But for the lack of real test datasets which could be publicly available by the cooperators, we can only still use the old datasets to present our improvements.The two specific points are as below:
One is that if a point in one sub-carrier array has been categorized as noise, then the sub-carrier would be wrongly recognized as two sub-carriers. Another is that when the two or more neighbouring sub-carriers are very close to each other, the FCN-based method cannot distinguish the demarcation points between them correctly, and this would cause severe fault and leak detections.
Second, we add the complexity comparisons between SCN and FCN methods in the revised manuscript, Table 3, in Section 4.4 (P11, 12). The FLOPs and network parameters of the SCN models are significantly larger than the FCN, but the inference time cost only increases by 52.2% when the model scale both 13. Because the proposed SCN method is an end-to-end deep learning-based network without other post-processing, while the FCN-based method only predicts the classification probability of each point of the inputs, they can not directly get the detection results. At the same time, the SCN can avoid the two problems mentioned above while FCN can not.
Point 5: More compassion results with other methods should be presented.
Response 5: Thank you for your comment. In Section 4.4 (P11, 12), we have added three other methods for the performance comparisons, including two threshold-based methods and one deep learning-based method, double-thresholds-based, slope tracing and SigdetNet, respectively.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors describe a novel neural network architecture intended for carrier frequency detection. Advantages of the article: 1) A novel neural network architecture - SCN. 2) The authors implemented SCN, trained and checked the performance. Disadvantages of the essay: 1) Without the author’s code, the reader can not replicate the article’s results. 2) The authors used a proprietary set of training data. It is challenging to compare systems without the same data. 3) The authors did not compare their results with others known from the literature.
Author Response
Response to Reviewer 2 Comments
The authors describe a novel neural network architecture intended for carrier frequency detection.
Advantages of the article:
1) A novel neural network architecture - SCN.
2) The authors implemented SCN, trained and checked the performance.
Point 1: Without the author's code, the reader can not replicate the article's results.
Response 1: We are very sorry for this. Because conflicts of interest and intellectual property issues exist with the cooperators, it is not proper to publish our code. We are very sorry again.
Point 2: The authors used a proprietary set of training data. It is challenging to compare systems without the same data..
Response 2: We greatly appreciate your comment. As the same as the point 1, it is very sorry for publishing our data. But people who are interesting in our work can send emails to us, and we would provide some training simulation signals to them.
Point 3: The authors did not compare their results with others known from the literature.
Response 3: Thank you for your comment. In Section 4.4 (P11, 12), we have added three other methods for the performance comparisons, including two threshold-based methods and one deep learning-based method, double-thresholds-based, slope tracing and SigdetNet, respectively.
Author Response File: Author Response.pdf
Reviewer 3 Report
The authors have studied an end-to-end deep convolutional neural network (CNN) model for carrier signal detection in the broadband power spectrum, named Spectrum-Center-Net (SCN). The topic is very interesting, however, reviewer has the following comments: 1) Motivation behind CNN model must be polished. Since deep neural networks (DNN) (for details refer Block deep neural network-based signal detector for generalized spatial modulation and A DNN Architecture for the Detection of Generalized Spatial Modulation Signals) may perform better than CNN. 2) Training parameters must be provided in a table for reader clarity. 3) What is time complexity of the proposed CNN detector? Moreover, compare it's time complexity and performance with conventional detectors such as ML and MMSE.
Author Response
Response to Reviewer 3 Comments
The authors have studied an end-to-end deep convolutional neural network (CNN) model for carrier signal detection in the broadband power spectrum, named Spectrum-Center-Net (SCN). The topic is very interesting, however, reviewer has the following comments:
Point 1: Motivation behind CNN model must be polished. Since deep neural networks (DNN) (for details refer Block deep neural network-based signal detector for generalized spatial modulation and A DNN Architecture for the Detection of Generalized Spatial Modulation Signals) may perform better than CNN.
Response 1: We appreciate your helpful comment. In our work, we aim to blindly detect the multi-signal exsiting in a wide spectrum band and estimate signal parameters such as the number of separable signals, centre frequencies, and bandwidths. But the two articles based on DNN mainly focus on Generalized Spatial Modulation (GSM) signals and the correctness of their bit imformation. Initially, we considered applying DNN to design a model, but for the long input length (32768), even if the three layers of the perceptron is difficult to train to converge for its high computation complexity and low feature extraction feature. But the CNNs weights sharing mechanism sharfly reduce the complexity of the model and can rapidly train to converge. So, at last, we give up the DNN-based methods.
Point 2: Training parameters must be provided in a table for reader clarity.
Response 2: We greatly appreciate your comment. We updated the manuscript and provide the training parameters in Table 1, Section 4.2 (P7). It is presented as below:
Table 1. SCN Training Parameters.
|
Implement Library |
PyTorch 1.10.0 |
|
Hardware Platform |
2 GeForce RTX 3080Ti GPU, Intel(R) Bronze 3204 CPU |
|
Operation System |
Ubuntu 20.04 |
|
Model Input Length |
32768 |
|
Batch Size |
32 |
|
Training Epochs |
150 |
|
Dropout Probability |
0.3 |
|
Optimizer |
Adam |
|
Learning Rate Strategy |
Cosine Annealing Warm Restarts, initial value 2e-5, T_0=10, T_mult=2 |
Point 3: What is time complexity of the proposed CNN detector? Moreover, compare it's time complexity and performance with conventional detectors such as ML and MMSE.
Response 3: Thank you very much for your comment. It is very true that we should compare the complexity of the proposed method and other existing methods. In Section 4.4 (P11, 12), we have added three other methods for the performance comparisons, including two threshold-based methods and one deep learning-based method, double-thresholds-based, slope tracing and SigdetNet, respectively. In addition, the ML and MMSE are improper in evaluating the multi-signal existing detection problem.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Authors have revised the paper based on my comments.
Reviewer 3 Report
No more comments from reviewer.
