Deep Learning Stranded Neural Network Model for the Detection of Sensory Triggered Events
Round 1
Reviewer 1 Report
1. The word "proposed" in the title has no importance and may be removed. In addition, it is preferable to avoid the use of abbreviations in the title. 2. Authors 1 and 2 have the same affiliation, so there is no need to write it twice. 3. Please avoid the use of short paragraphs in the introduction and everywhere else. 4. It is preferable to split the introduction into two sections: Introduction + Related Works. 5. The findings about existing related works may be summarized in tabular form. 6. The authors are invited to add a short paragraph about the use of formal techniques for checking the correctness of AI-based solutions. 7. For this purpose, the following references may be included: a. https://ieeexplore.ieee.org/document/9842406 b https://dl.acm.org/doi/abs/10.1145/3503914 8. The authors are invited to summarize their contributions in the introduction. 9. The paper contains too many tables. Some of them may be moved to the appendix. 10. The authors need to identify the shortcomings of their work and propose appropriate future word directions.Author Response
Thank you very much for your time and effort spent reviewing our manuscript. Here we quote our responses and amendments performed based on your comments.
Comment 1. The word "proposed" in the title has no importance and may be removed. In addition, it is preferable to avoid the use of abbreviations in the title.
Response 1: The word proposed has been removed from the title. As correctly mentioned, the NN abbreviation in the title has also been removed.
Comment 2: Authors 1 and 2 have the same affiliation, so there is no need to write it twice.
Response 2: Authors 1 and 2 affiliations have been merged.
Comment 3: Please avoid the use of short paragraphs in the Introduction and everywhere else.
Response 3: Short paragraphs in the Introduction have been joined, maintaining the meaning. The same applied to the Materials and methods and, experimentation, discussion of the results sections.
Comment 4: It is preferable to split the Introduction into two sections: Introduction + Related Work.
Response 4: The introduction section has been split into Introduction and Related Work sections.
Comment 5. The findings about existing related works may be summarized in tabular form.
Response 5: The findings about existing related works have been arranged into two subsections. The part of critical events detection and periodic Industrial maintenance tasks part. Summarization per part (subsection) has been performed using three axis: 1. Existing threshold-based, Fuzzy, and PID methods (first 1-2 paragraphs), 2. Conventional Machine Learning methods (1-2 paragraphs that follow) and 3. Deep learning methods (last paragraphs).
Comment 6: The authors are invited to add a short paragraph about the use of formal techniques for checking the correctness of AI-based solutions. For this purpose, the following references may be included: a. https://ieeexplore.ieee.org/document/9842406 b https://dl.acm.org/doi/abs/10.1145/3503914
Response 6: Appropriate paragraph has been added to the Introduction section (4th paragraph -line 46), including the mentioned references.
Comment 7: The authors are invited to summarize their contributions in the Introduction.
Response 7: The last paragraphs of the original Introduction section summarizing the authors' contributions have been placed into the new Introduction section as one penultimate paragraph. The last paragraph in the Introduction section describes the structure of our manuscript.
Comment 8: The paper contains too many tables. Some of them may be moved to the appendix.
Response 8: Cross comparison Tables of the size and trainable parameters of LSTM and Strandded-NN models have been moved to the Appendix section.
Comment 9: The authors need to identify the shortcomings of their work and propose appropriate future word directions.
Response 9: Shortcomings of our work have been amended and added in the conclusions section. Future work directions have been added in the last paragraph of our conclusions section.
Reviewer 2 Report
The authors proposed a stranded -NN method to process various data for predictive maintenance. The paper is logically organized and easy to follow. The main concern is regarding novelty and significance of the proposed method:
1. the authors proposed NN with different layers/neurons using Eq.1,2, etc. Thresholds were given without justification, which means that the given thresholds may work well for the experimental data used in the manuscript. But for readers with their own data, would these thresholds still work well? Thus, applications of the proposed method/thresholds are limited.
2. if the readers need to adjust the method/thresholds based on their previous experience, the proposed method actually didnot help readers very much because they may have used NNs with different layers/neurons already just to try and error in order to get better results.
3. no new knowledge was generated either in thresholds selection or NN structural design, thus scientific contributions of the manuscript is limited.
some typos, such as
- at line 298, where "wither" should be replaced with "either."
Author Response
Thank you very much for your time and effort spent reviewing our manuscript. Here we quote our responses and amendments performed based on your comments.
Comment 1. The authors proposed a stranded -NN method to process various data for predictive maintenance. The paper is logically organized and easy to follow. The main concern is regarding novelty and significance of the proposed method:
Response 1: The novelty of our proposition is the stranded-NN algorithm can apply to periodic (periodic maintenance detection), close to real-time and real-time events (operational events detection) as a unified entity. It can maintain similar accuracies with LSTM implementations and train faster than LSTM implementations that still maintain a fixed number of cells (long memory). If keeping the long memory long, then training time increases exponentially concerning the stranded-NN model. Furthermore, in cases where the LSTM may have better accuracy results. It can also be attached to a strand in our model for specific input batch size (time interval x sensory measurements). As mentioned by the authors' such cases are the cases of predictive maintenance of more than hourly or even daily interval size steps. Appropriate amendments have been made in the discussion and conclusions sections so as emphasize on the novelty and significance of our proposition
Comment 2: the authors proposed NN with different layers/neurons using Eq.1,2, etc. Thresholds were given without justification, which means that the given thresholds may work well for the experimental data used in the manuscript. But for readers with their own data, would these thresholds still work well? Thus, applications of the proposed method/thresholds are limited.
Response 2: Appropriate justification of the m threshold value has been added to the paragraph at line 291. Also, example cases of model strand generation have been described for real-time and close-to-real-time detections.
Comment C3:. if the readers need to adjust the method/thresholds based on their previous experience, the proposed method actually didnot help readers very much because they may have used NNs with different layers/neurons already just to try and error in order to get better results.. no new knowledge was generated either in thresholds selection or NN structural design, thus scientific contributions of the manuscript is limited.
Response C3: The proposed stranded-NN algorithm's variable batch data input (step of measurements) signifies the time depth of monitoring and therefore expresses the sensitivity level required for each requested maintenance process accordingly. For example. If someone monitors a set of N temperature sensors that transmit measurements every minute, then by having a NN that has as input 1-hour measurements, meaning 60xN input in its NN network, It performs a macroscopic detection for abnormalities close to predictive maintenance checks. If, however, training a NN that has real-time inputs is much more susceptible to change and enforces a sensitive real-time detector-classifier for critical machinery parts. Both these capabilities can be unified with the stranded-NN algorithm to a set of NN model strands that respond as a single entity – entry point and activate different neural networks for different detection approaches based on the number of sensory inputs (and over time) fed to the stranded-NN model as a batch input 1D array.
Appropriate amendments have been made in the results and discussion section so as to clearly present the authors' stranded-NN model advantages over MLP and LSTM models.
Comment 4: some typos, such as
- at line 298, where "wither" should be replaced with "either."
Response 4: The paragraph at line 298 has been rewritten. The manuscript has been thoroughly checked for grammatical and typo errors
Reviewer 3 Report
The paper faces the use of decision making Machine Learning (ML) techniques in order to improve the robustness in industrial maintenance, detecting the majority of possible fault through pattern recognition, and triggering a proper alert. The paper proposes an intelligent failure classification algorithm called the stranded-NN model which is used to detect different classes of Industrial emergencies based on input time-depth of sensory measurements.
The authors compared the classification accuracy of their model with existing models, such as deep MLPs and LSTMs of different cell sizes, for real-time and near-real-time classification cases of compressor temperatures and pump instabilities recorded as acceleration.
From the authors' experimentation, the Stranded-NN significantly outperformed its MLP and LSTM counterparts in near real-time classification.
Some main questions to address:
• In the section 2.4. Proposed Stranded-NN model the authors provide both methods to adjust the number of model strand depth of hidden layers q and number of trainable parameters of the model in equations 1 to 3, but some discussion about the origin and motivation of these equations will be of help to understand the proposal.
What are the hypothesis behind the ecuations?
Why are choosen the thresholds for m (2 ≤ m ≤ 32 and m>32 )?
Any other value for the thresholds have been validated during the research? If yes, what were the results?
Some minor aspects to be addressed are the following:
• The authors could include a citation regarding the paragraph on line 191
Some typos:
• Line 227, "Industrial"
Author Response
Thank you very much for your time and effort spent reviewing our manuscript. Here we quote our responses and amendments performed based on your comments.
Comment 1: In the section 2.4. Proposed Stranded-NN model the authors provide both methods to adjust the number of model strand depth of hidden layers q and number of trainable parameters of the model in equations 1 to 3, but some discussion about the origin and motivation of these equations will be of help to understand the proposal.
What are the hypothesis behind the equations?
Why are choosen the thresholds for m (2 ≤ m ≤ 32 and m>32 )?
Any other value for the thresholds have been validated during the research? If yes, what were the results?
Response 1: Appropriate justification of the m threshold value has been added to the paragraph at line 291. Also, example cases of model strand generation have been described for real-time and close-to-real-time detections.
Some minor aspects to be addressed are the following:
Comment 2: The authors could include a citation regarding the paragraph on line 191
Response 2: paragraph on line 191 has been revised, and references have been added.
Comment 3: Some typos:
- Line 227, "Industrial"
Response 3: Industrial commentators have been amended to concentrators. The manuscript has been thoroughly checked for grammatical and typo errors
Round 2
Reviewer 1 Report
The authors considered my comments and suggestions. Good luck.
