Deep Neural Network for Valve Fault Diagnosis Integrating Multivariate Time-Series Sensor Data
Round 1
Reviewer 1 Report
Comments and Suggestions for AuthorsThis paper represents a Deep Neural Network solution for valve diagnosis using multivariate time-series sensor data. The topic is interesting, but the paper needs improvement before the publication. Please pay attention to the following comments:
The abstract needs re-writing for clarity and to show the motivation of the research
There is no literature review about the fault diagnosis of valves, as most of the literature is generic and not related to the selected component(valve)
The literature review should cover the limitations of data drive approaches, such as the availability of the data and the ability to capture all fault scenarios.
The authors didn’t use the correct terminology in the paper. They referred to valve failure as it should be valve failure. There is a big difference between the two, as failure is very easy to detect, while the early detection of the fault is a big challenge. Also, they used water quantity, which should be the water flow rate. In addition, he is using the damage score instead of the severity of the fault.
The experiment diagram is flawed because it doesn’t show the main tank, the sump tank, or how the water goes from the nozzle valves to the tank.
The experiment is over-instrumented with sensors, so a sensor optimization study is required to identify the number and location of the sensors.
Using a filter at the inlet pipe is strongly recommended whenever you have circulated fluid to avoid big foreign objects in the circulation.
The type of valves should be specified because they are equipped with motors that control their opening and closing levels,
The title of Figure 3 is incorrect.
The samples of collected data are not represented in the paper.
There is no data processing, as mentioned in the paper.
It’s not clear how the authors emulated the fault in the experiment. They stated, "In this experiment, a significant pressure drop was observed at the valve outlet where a foreign object was inserted, with an internal diameter of 3.0 mm.”. Then, they stated that the valves have motors controlling their opening and closing levels.
The authors represented the fault severity by changing the inner diameter of the pipe as a linear relationship. However, this relationship is not linear.
The authors stated, “To test the fault diagnosis capability on valve types not used during the training phase,..”. So, why are you testing the algorithm on different valve types?
The Heatmap visualisation of the dataset in Figure 4 does not represent the values of pressure and flow rate in the experiment. It does not reflect the magnitude of change in each scenario.
It’s unclear how the motor angle change represents different levels of fault severity and how the internal diameter of the valve will change.
Why the start value of the non-operating valve angle is different between the training and testing datasets?
Why are the samples of the testing dataset more than the training dataset? Generally, the ratio is 20% to 80% or 30% to 70%.
Extensive formatting and proofreading are needed. For example, the text in lines 252-257 is light grey.
The authors stated, “The 'Only types used when testing' row represents results when testing only 'unseen' failure types that were not used during training. Although achieving lower accuracy than the trained types, it still demonstrates significant results with an MAE of 0.0467 and an RMSE of 0.0954.”. It is wrong because the MAE and RMSE of trained data are 0.0024 and 0.0059, respectively.
The proposed solution does not diagnose unseen failures that do not exist in the training data. It just diagnoses different levels of fault severity.
In general, the paper needs restructuring to be comprehensive.
Comments on the Quality of English Language
Extensive formatting and proofreading are needed.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for Authors- How robust is the model to noise and variations in real-world operational conditions beyond the experimental setup?
- What is the computational cost of the proposed network, and how scalable is it to larger systems?
- Were any specific pre-processing techniques used for the time-series data, and how did they impact performance?
- How does the model's performance compare to simpler, potentially more interpretable methods?
- What are the limitations of the model, and what future work is needed to address them?
- How generalizable is the model to different valve types, fluids, and system architectures?
- It is suggested to go through the following papers:
· Advancing machine fault diagnosis: A detailed examination of convolutional neural networks
· A roadmap to fault diagnosis of industrial machines via machine learning: A brief review
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Comments and Suggestions for AuthorsThere is a significant improvement compared to the previous version. I suggest addressing the following points before publication:
- Rewrite the abstract to enhance its structure, clarity and contribution.
- The pressure sensor readings in Fig. 5 are very noisy, especially sensors 11, 12 and 13. Can you explain how these noisy data can be used directly as input to the algorithm without processing, as stated in lines 266-268?
Comments on the Quality of English LanguageThere is a need for proofreading before the publication
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 2 Report
Comments and Suggestions for AuthorsNo Further Comments
Author Response
We sincerely appreciate the valuable insights and constructive feedback you have provided throughout the review process. Your comments have been instrumental in improving the quality and depth of our paper. We are truly grateful for your time and effort in reviewing our work.
Thank you once again for your thoughtful contributions.