Intelligent Fault Diagnosis for Inertial Measurement Unit through Deep Residual Convolutional Neural Network and Short-Time Fourier Transform
Round 1
Reviewer 1 Report
This paper is a great effort of the authors. However, I appreciate a few more improvements:
· The problem statement needed to be more precise in the abstract.
· The contribution part in the Introduction shall be well identified and presented in a specific manner.
· In the introduction section, a few more literatures can be great for the authors. As an example, for STFT-based analysis I recommend including https://doi.org/10.1016/j.measurement.2020.108478 this one.
· Figure 9 needs to be revised. There is a huge problem with it. This is not from CWRU.
· Future directions should be well highlighted in the conclusion.
Author Response
Please see the attachment.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper is very interesting and solves, with excellent performance, a complex fault diagnosis problem from the Inertial Measurement Unit data. All of this process is based on a custom deep learning neural network.
Please correct the following issues:
· In lines 50-51-52, you make the following affirmation: “traditional machine learning based fault classification methods show much limitations and poor performance in term of processing speed”. The processing speed is mainly a problem of the deep learning structures, not of the classical learning machine algorithms.
· In relation (8), you have h (m-n), not h (n-m)
· The same problem is presented in relation (9): h (m-n)
· In Figure 5, you have two consecutive Samples 5 segments
· Please explain more clearly the difference between “the length of window is set to 64” (in lines 270 and 271) and “small slices with the length of 1024” (line 286)
· Where are in Figure 7 the “adaptive maximum pooling layer”? The „maximum pooling layers” is present, but the “adaptive maximum pooling layer” is not depicted.
· Leave an empty line after each figure and before each table.
· In the text description for Figure 9, what is the CWRU?
· How do you transform the 33 x 33 spectrum images to 330 x 330? Please give the readers more information.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Reviewer 3 Report
Reviews for machines
This manuscript proposes a deep learning-based approach for fault diagnosis (FD). In which, STFT is used to obtain the time-frequency expression of raw signals and is fed into a deep residual convolutional neural network to extract hidden features for FD. This paper is well organized, the methodology is somehow novel, and the experiments on two public data sets confirmed its effectiveness. However, the reviewer still has some concerns, as follows:
1) The authors should check the grammar again; some sentences are hard to read. E.g., in lines 145-146, what is else improved method should be given?
2) Why did you apply STFT to get a time-frequency expression of raw signals that should be given? Why not use other frequency-domain transformation methods?
3) Similar to issue 2, why applied Z-score normalization, not others?
4) This article is based on the CNN approach, but the reference for the CNN-based method is not enough. Only one is related to the CNN-based approach. The following good references should be considered: https://doi.org/10.3390/s22114156 and doi: 10.1109/TIE.2018.2844805.
5) Formula 1 is incorrect as convolution operation is not included in your given expression.
6) One ablation study should be given to explore each component’s effectiveness in the proposed method since your method is combined serval components.
Author Response
Please see the attachment.
Author Response File: Author Response.pdf
Round 2
Reviewer 1 Report
Congratulations.
Reviewer 3 Report
Thanks for the authors' careful response and revision, this version looks better. It solved all my concerns and I think it can be published in the current form.
