Fault Diagnosis in Drones via Multiverse Augmented Extreme Recurrent Expansion of Acoustic Emissions with Uncertainty Bayesian Optimisation

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

This manuscript presents a multiverse augmented recurrent expansion method for fault diagnosis of drones. This is an important issue in the field. Some comments are listed as follows.

1. It is suggested to clearly explain the use of acoustic emission data in the diagnosis of drones. What are the diagnosis tasks in drones by using such data?

2. The authors should explain the domain-specific issues of using data-driven methods. The listed issues, such as variability, noise, and outliers, are general in the field. More descriptions about the reason of occurring data problems need to be added.

3. The contributions need to be further refined. It is suggested the authors summarize the highlights of this submission but not show the method steps. Some contributions, such as evaluation metrics and evaluation datasets, do not refer to the innovations.

4. The authors should give the objective functions of the proposed method with a connection to Figure 5. It is also necessary to show the training process of the proposed method.

5. Data decentralization and label noise are important issues in data-driven diagnosis methods. Interestingly, the authors could discuss these topics in the manuscript.

- Targeted transfer learning through distribution barycenter medium for intelligent fault diagnosis of machines with data decentralization

- Label recovery and trajectory designable network for transfer fault diagnosis of machines with incorrect annotation

Author Response

Dear Associate Editor and Reviewers,

The authors are thankful for the efforts of the Editor and the Reviewers for the evaluation of our proposal. The relevant comments and suggestions have definitely improved both the proposal presentation and quality. In this context, the authors have incorporated all the suggestions in the revised manuscript and have addressed all the raised issues. In this document, we propose detailed responses to the comments and questions. All the revisions are highlighted in red in the revised manuscript.

Author Response File: Author Response.pdf

Reviewer 2 Report

Comments and Suggestions for Authors

Good manuscript! Comprehensively introduced, easy to follow, well written. Still, to fully satisfy readers tempted to apply the newly proposed efficient learning method for their similar problems at hand, some issues remain.

Major issues:

1) Abstract: To appreciate to impressive results (e.g. >99 % accuracy of classification and hence more than 8 % higher compared to LSTM while up to 10 times faster) and draw attention of the quick reader, please include numerical results, both of accuracy and improvement to known methods (LSTM).

2) Line 200: What do you mean by "drone signals"? What are the lengths (min, max, mean, std) of these sound fragments? What are the proportions of N, MF, PC signals?

3) Instead of writing "progressive increase in variability, amplitude, and fluctuations over time", please give numbers of the parameters which are distinctive for the datasets A, B, C. This clearly would help to appreciate the results in Tables 1-4. Are all 3 drone types (also named A, B, C) involved? Is it consistent, that the coefficient of determination (Fig. 6) is higher for increasing variability and fluctuations (datasets B and C compared to dataset A)?

4) Fig. 3. What is the reason for sequencing the sound fragments of different origins in this specific order? Would changing the sequencing order effect the results? Why is the fragment of normal operation much smaller than that of malfunctions? How would the length of the data affect the results? Is there a minimum length of the recorded sound data to make the algorithm work properly?

5) Figure caption of Fig. 3. For clarification, please add to the figure caption the color coding, i.e. the normal operating mode (N) and two types of malfunctions, motor fault (MF) and propeller cut (PC). Please make clear in the figure caption, what are the characteristic values of datasets A, B, C.

6) Fig. 7: The purpose and meaning of the figure is difficult to understand. The figure caption is cryptic.

7) According to Fig. 8, there were ~300 class representatives in the 3 datasets. Does this mean, that there were 54.000 / 300 = ~180 sound samples per class representative?

8) In order to avoid toggling between Table 2 and Table 4, please consider combining these results in one Table (adding the training/testing time to the UBO-MVA-EREX method. This way, the tremendous improvements in performance metrics and computational time would be clearly visible.

9) Please improve the Conclusions. Be more specific, e.g. what "refining the methods" (Line 547) or "notable enhancements across various error metrics, classification metrics, uncertainty metrics, and computational time, both visually (??) and numerically (Line 538f.) would mean in numbers. What do you mean by "particularly as dataset complexity escalates"? (Line 539). If other malfunctions would be worth consideration for drone fault detection,  please state so and make some comments.

Minor issues:

1) Abstract, Line 14f: "data representativeness methods, such as learning systems" is unclear

2) Abstract, Line 18: instead of "interesting representation learning philosophy" it is advised to write "efficient representation learning method".

3) Abstract, Line 29f. Check language and consider rearranging the fragment "of the proposed scheme for such challenging tasks of fault diagnosis based acoustic emission of drones."

Comments on the Quality of English Language

Text is fluently readable, use of English language is mostly correct. Some issues were found and indicated to the Authors.

Author Response

Dear Associate Editor and Reviewers,

The authors are thankful for the efforts of the Editor and the Reviewers for the evaluation of our proposal. The relevant comments and suggestions have definitely improved both the proposal presentation and quality. In this context, the authors have incorporated all the suggestions in the revised manuscript and have addressed all the raised issues. In this document, we propose detailed responses to the comments and questions. All the revisions are highlighted in red in the revised manuscript.

Author Response File: Author Response.pdf

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

The authors have well addressed all my comments, and it could be accepted now.