A Fiber Vibration Signal Recognition Method Based on CNN-CBAM-LSTM

Round 1

Reviewer 1 Report

The authors have done an interesting work on enhancing signal recognition based on CNN-CBAM-LSTM. The goal of this work is the optimization method to minimize the error recognition.

The manuscript can be considered for publication after addressing the following comments:

1.  The neural networks and different types of artificial intelligence are very interesting as application method to enhance various aspects of CNC field. For these reasons introduction section should be implemented and you can cite the papers:

- Xin, R.; Zhang, J.; Shao, Y. Complex network classification with convolutional neural network. Tsinghua Sci. Technol. 2020, 25, 447–457, doi:10.26599/TST.2019.9010055.

-  Kattenborn, T., Leitloff, J., Schiefer, F., & Hinz, S. (2021). Review on Convolutional Neural Networks (CNN) in vegetation remote sensing. ISPRS Journal of Photogrammetry and Remote Sensing, 173, 24-49.

- Bhatt, D., Patel, C., Talsania, H., Patel, J., Vaghela, R., Pandya, S., ... & Ghayvat, H. (2021). CNN variants for computer vision: History, architecture, application, challenges and future scope. Electronics, 10(20), 2470.

- Baiocco, G.; Almonti, D.; Genna, S.; Ponticelli, G.S.; Tagliaferri, V.; Ucciardello, N. Neural network implementation for the prediction of load curves of a flat head indenter on hot aluminum alloy. Procedia CIRP 2020, 88, 543–548, doi:10.1016/j.procir.2020.05.094.

-  Baiocco, G.; Almonti, D.; Guarino, S.; Tagliaferri, F.; Tagliaferri, V.; Ucciardello, N. Image-based system and artificial neural network to automate a quality control system for cherries pitting process. Procedia CIRP 2020, 88, 527–532, doi:10.1016/j.procir.2020.05.091.

- Xie, X., Cheng, G., Wang, J., Yao, X., & Han, J. (2021). Oriented R-CNN for object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 3520-3529).

- Almonti, D., Baiocco, G., & Ucciardello, N. (2021). Pulp and paper characterization by means of artificial neural networks for effluent solid waste minimization—A case study. Journal of Process Control, 105, 283-291.

 

2.  The authors should check all the acronymises and add long version first appearing in the manuscript;

3.  In the 5.3 CNN-CBAM-LSTM Model Introduction section, the authors should realize a workflow scheme to increase readability of process;

4.  Figure 6 shows low quality, the authors should change itself with the same but higher quality;

5.  Figure 7 shows low quality, the authors should change itself with the same but higher quality;

6.  In table 3, OFVS results are missed, in the previous text, the authors said that in table 3 there is a comparison but the main results missed. The authors should add OFVS data;

7.  The analysis of this kind of signal is fundamental in relation of processing time. The authors should show the obtained improvement results of their method in comparison with other methods;

 

8.  The results show high accuracy, the authors should discuss the signals that the network cannot detect. Are there any particular waveforms? Does changing the training and verification dataset change anything? 

Author Response

Thank you very much for your suggestions on my paper. Here are my answers to your questions:

1. Thank you for your suggestion. I have read your recommended paper and quoted it in my introduction section. Please refer to the paper for details.

2. Thank you for your advice. I have checked my paper and found several problems of improper application of abbreviations, which have been corrected. Thank you again.

3. Thanks for your suggestion. I have added a workflow scheme in section 5.3 of the paper. Please refer to the paper for details.

4. Thanks for your advice. I have improved the clarity of Figure 6.

5. Thanks for your advice. I have improved the clarity of Figure 7.

6. Sorry, maybe the description in my paper is not clear. Here is the explanation: The Optical Fiber Vibration Signal (OFVS) used in this paper contains four different types: Flap, Knock, Walk and Run. In Table 3, VMD, AFCNN, CNN-LSTM, and CNN-CBAM-LSTM are respectively used to realize the pattern recognition processing of OFVS, and the accuracy of the four methods is compared, so as to get our conclusion.

7. Sorry, we do not quite understand The sentence "The analysis of this kind of signal is fundamental in relation to processing time". Can you describe the problem in detail and what part of the paper it relates to? Thank you very much.

   I will try to answer the second part of your question as I understand it:

   Comparing the proposed method with the traditional method VMD, it can be found that the CNN-CBAM-LSTM way has significantly improved the recognition effect of the four types of OFVS, and the model parameters can be automatically updated through network iteration, avoiding the limitation of manual parameter design of the traditional method. Compared with AFCNN and CNN-LSTM methods, it can be seen that our proposed method has higher average accuracy and better network fitting effect.

8. I will answer your questions separately:

   Answer 1: The actual situation of our study is not a complex environment in which noise swamps signals. In actual situations, useful signals and noisy signals can be distinguished by technical methods. So, with the current device, we don't detect a situation where the signal is drowned in noise, so there's no situation where we can't detect the signal.

   Answer 2: There are no other special waveforms in our signal.

   Answer 3: In Section 5.2 Dataset Processing, we introduced the Processing method of the Dataset. In order to reduce the error caused by the singleness of the training set and validation set, we used the hold-out method to divide the Dataset, randomly selecting 80% of the total amount as the training set and 20% as the test set each time, and finally taking the average value after repeated experiments. Therefore, the training set and test set of each experiment will be changed randomly. The experimental results show that if the training set and test set are changed, there may be some minor fluctuations in the results, which will not have a great impact on the experimental results.

Reviewer 2 Report

 

The paper shows a CNN-CBAM-LSTM network model that recognizes the fiber vibration signal. The authors demonstrated that this method is better than VMD, AF-CNN, CNN-LSTM.

The authors should read the paper carefully to improve the quality of the English and the ease of comprehension. Below are some examples:

-       Avoid repetition (line 18, 195)

-       Rephrasing (line 42, 60)

-       Typos (line 69)

 Minor comments:

-       Improve the quality of figures 6 and 7

-       Table 1 is split into two pages

-    Position the reference correctly and not in the middle of the sentence

Author Response

Thank you very much for your suggestions on my paper. Here are my answers to your questions:

1. Thank you for your advice. I will read the article carefully to improve my English ability. I checked and revised the paper, hoping to reduce the problems in grammar and writing.

Thank you for listing some examples, such as Avoid repetition (lines 18, 195), Rephrasing (lines 42, 60), and Typos (line 69), I have made some modifications to the paper.

2. Thanks for your advice. I have improved the quality of Figure 6 and Figure 7 (now labeled Figure 9 and Figure 10) in my paper.

3. Thanks for your suggestion. I have improved the problem of "Table 1 is split into two pages" in my paper.

4. Thanks for your suggestion, I have changed the position of the reference in the sentence in the paper.

Reviewer 3 Report

-The paper should be interesting ;;;

-it is a good idea to add a block diagram of the proposed research (step by step);;;;;;

-it is a good idea to add more photos of measurements, sensors + arrows/labels what is what (If any);;;

-What is the result of the analysis?;;

-figures should have high quality;;; 

-labels of figures should be bigger;;;;

-Fonts of figures should be corrected;;;

-In figure 4 please add SI units to Amplitude;;;

-please add photos of the application of the proposed research, 2-3 photos (if any) ;;; 

-what will society have from the paper?;;

-Please compare the proposed method with other approaches;;

-Is there a possibility to use your approach/research for other problems of fault diagnosis;

-references should be from the web of science 2020-2022 (50% of all references, 30 references at least);;;

-Conclusion: point out what have you done;;;;

-please add some sentences about future work;;;

Author Response

Thank you for your suggestions on my paper. Here are my answers to your questions:

1. Thank you for your advice. I will try to modify some expressions to make the article as interesting as possible.

2. Thanks for your suggestion. I have added a block diagram of research steps in section 5.1 of the paper. Please refer to the paper for details.

3. Thank you for your suggestion. I have added some photos of the actual measurement environment (including two types of hanging net and buried) in section 5.1 of the paper. Please refer to the paper for details.

4. Through our experimental results, it can be seen that the CNN-CBAM-LSTM model proposed in this paper has a good effect on the recognition of fiber vibration signals. The model can be updated automatically through network iteration, which has better generalization ability than the traditional pattern recognition algorithm. Compared with other neural network models (such as AFCNN, CNN-LSTM, etc.), the accuracy is also improved.

5. Thanks for your suggestion. I have modified the clarity of Figures 6 and 7 (now labeled as Figure 9 and Figure 10) in the paper.

6.  Thanks for your suggestion. I have modified the label size of some figures on the paper.
7. Thanks for your advice. I have corrected the fonts of some figures in my paper.

8. Thank you for your suggestion. The problem that SI units are not added in Figure 4 is explained as follows: Figure 4 shows the results of Signal Preprocessing of four types of OFVS. We introduced the basic process of Signal Preprocessing in Section 5.1 Signal Acquisition and Preprocessing. Among them, there is a step to maximum normalized signal processing (which can also be called a feature normalization), numerical size range is not possible due to the various characters, through the characteristics of the normalized value range, can eliminate dimensions between different samples, orders of magnitude of the difference of characteristic properties such as the characteristics of comparable, convenient behind the study of signal and processing. Therefore, after the maximum normalization operation, the amplitude value of the signal is transformed into a dimensionless relative value in the [-1,1] range, so SI units are not added to the amplitude in Figure 4.

9. I'm sorry, we don't have any photos of the research projects we're applying for.

10. With the development of society, distributed fiber perimeter security system has been widely used in various important areas, and the correct identification of intrusion behavior has become very important. The method proposed in this paper can be used in perimeter security systems using fiber optic sensors, which contributes to the accuracy of security system identification and reduces the loss caused by false alarms.

11. Compared with the traditional VMD method, the CNN-CBAM-LSTM method proposed in this paper significantly improves the recognition accuracy of OFVS and does not rely on manual design parameters, with better generalization ability. Compared with deep learning methods such as AFCNN and CNN-LSTM, it can be seen that our proposed method has higher average recognition accuracy, but the recognition rate of running signals is relatively low.

12. The fiber vibration signal we study is nonlinear, which is different from the common sound vibration signal. Therefore, our research can be used in some cases, such as vibration signal detection of transformers in a harsh electromagnetic field environment, intrusion detection of some oil pipelines, etc. However, our method cannot be used to detect weak vibration signals.

13. Thanks for your advice. I have revised the reference section of the paper. Please refer to the paper for details.

14. Thank you for your suggestions. I have updated my main work in the Conclusion of the paper. Please refer to the paper for details. Here is a brief overview:

    In terms of data processing, the spectral subtraction method is used for denoising, and the spectral entropy method combined with short-term logarithmic energy is used for endpoint detection, and the starting position is found to intercept the isometric signal and finally normalized.

    In terms of OFVS signal recognition, this paper constructs a deep learning network model of CNN-CBRAM-LSTM, which performs feature extraction, feature refinement, feature learning, and other processing on the signal, and finally completes the signal recognition task combined with MLP. In the experimental part, the model is trained and tested, and the conclusion that the proposed method is feasible and effective for the identification of fiber vibration signals is drawn by comparison.

15. Thank you for your suggestion, a brief sentence about future work has been added to the paper:

    In the future, with the continuous development of artificial intelligence algorithms, we can continue to study more excellent OFVS recognition methods from the aspects of endpoint detection and recognition algorithms.

Reviewer 4 Report

Article “A Fiber Vibration Signal Recognition Method Based on CNN-CBAM-LSTM” is devoted to the development of  a method for processing signals from vibrations arriving at an optical fiber, with the involvement of convolutional neural network.  Article is of important practical importance and corresponds to the theme of the journal Applied Science.

Nonetheless, to improve the article, I recommend that the authors pay attention to the following comments:

 

1. The authors should indicate more clearly in the introduction the specific purpose of the work.

2. I would recommend that authors present mathematical equations not in plain text, but with the help of special applications, such as Equation.

3. The quality of the figures, especially 6, could be improved

4. As far as I understand, OFVS is translated as optical fiber vibration signal. However, I did not find a direct indication of this in the article. I would ask the authors to clarify this in the manuscript.

I believe that after improvement, the article can be published.

Author Response

Thank you for your advice, and here are my answers to your questions:

1. Thanks for your advice. I added the content about the specific purpose of my research in the introduction of the paper. Please refer to the paper for details.

2. Thank you very much for your advice, I would like to make some explanations for this problem: I used the formula editor that came with Microsoft Office Word version 2019 when writing mathematical formulas. Later, I changed the font of the formulas to be the same as the font of the paper template, so it may look like plain text.

3. Thanks for your advice, I improved the quality of Figures 6 and 7 (now labeled Figures 9 and 10) in the paper.

4. Thank you very much for your suggestion. OFVS stands for fiber optic vibration signal. Sorry, it was an oversight on my part not to announce it when it first appeared. I have revised the paper. Thanks again for your advice.

Round 2

Reviewer 1 Report

The authors improved manuscript in accordance with reviewers' suggestions. The manuscript can be considered for publication in this new form.

Author Response

Thank you very much for your valuable suggestions on our paper, and thank you for your approval of our revised paper.

Reviewer 3 Report

-Formula for accuracy should be added;;;

-figures should have good quality;;;

-SI units should be added for example AMplitude of what [m/s^2] or what?

-Fig. 3 add arrows what is what;;;;

Author Response

Thank you very much for your advice. Here are my answers to your questions:

1. Thank you very much for your advice. The accuracy formula (Formula 10) has been added to the paper.

2. Thanks for your advice, I've added comments to Figure 3 to make it more readable. However, due to the limitations of the shooting equipment, the definition of the photo is not very high. I am very sorry for this.

3. Thank you for your advice. Sorry, my previous explanation about this problem is not clear, I will explain it again here.

   As different invasion behavior occurs, the fiber Sagnac interferometer can be detected by the different phase difference of the variation in light intensity, and then through the role of the photodiode will luminous power into electrical current, electricity flows through trans-impedance amplifiers into voltage, voltage is amplified and then through the role of ADC (analog-to-digital converter) into a digital signal. In the whole process, due to the different vibration sources and photoelectric converters at the receiving end at different locations, the transmission distance and the number of components passing through in the transmission process are also different, leading to different transmission losses of signals collected at different detection points during the whole transmission process. Therefore, it is easy to see that the amplitude of OFVS of the same type varies greatly due to different detection locations. In addition, when we do pattern recognition work, there is no great relationship between the signal's amplitude and the signal's type. The absolute magnitude of the amplitude has no practical significance for pattern recognition, so we only need to know the relative quantity obtained after normalization. The spectral structure of the signal is more important than the amplitude.

   In the preprocessing of the signal, we add a part of normalization processing, which adjusts the data according to the maximum value in a set of data, so that the signal's amplitude is transformed into a dimensionless relative value in the range of [-1,1]. It has the following advantages: on the one hand, it can reduce the negative impact of the amplitude difference caused by transmission loss on the subsequent feature extraction; On the other hand, by normalizing the value range of features, the difference of orders of magnitude between different samples can be eliminated to make features comparable. If the unnormalized data are modeled directly, the neural network model may learn too many variables with large values and ignore some important features, resulting in insufficient model training.

   To sum up, it is effective and necessary for us to normalize the signal. After normalization, the signal in FIG. 6 has become dimensionless data, so the amplitude in the figure cannot be expressed in SI units.

4. Thanks for your suggestion, we have added arrows and instructions in Figure 3.