Multi-Scale Recursive Semi-Supervised Deep Learning Fault Diagnosis Method with Attention Gate

Round 1

Reviewer 1 Report

The paper introduces a novel method for semi-supervised fault detection using deep learning. The authors compare it to the current state-of-the-art methods and show an improvement in fault detection and classification. The main novelty of the model is the introduction of an attention gate and a multi-scale feature reconstruction (which if I understand correctly is adding a higher level of the autoencoder to the training method). I would think that all the information will be encoded in the most abstract encoder layer, but it does seem to improve the results (and as the author cites there is current literature on the validity of the method).

While the improvement in accuracy is encouraging, it would've been nice to see a comparison using the full dataset as labeled data (to get a baseline on how good the model can be).

The authors also state in the conclusion that current methods cannot obtain satisfying results. While up to 7% improvement on the current method is an encouraging achievement it does not improve on the current state of the art in a sufficient way to completely distinguish the propsed model.

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 2 Report

This paper proposed a semi-supervised deep learning fault diagnosis model using attention mechanism to fuse multi-scale features and applied it to the fault diagnosis of rolling bearings. Some problems and suggestions are as follows:

1. One of the significant shortcomings of the Deep Neural Network is the expansion of the number of parameters caused by fully connected neurons. Under this condition, whether or not the introduction of the attention mechanism and longer memory loss function will lead to the reduction of the model training speed, and make the model more likely to fall into the overfitting state?

2. The author uses the CWRU experimental data set of Western Reserve University to verify the effectiveness of the proposed method. The vibration signal of the data set can almost be considered as the bearing state signal in a noise-free environment, but the vibration signal of the bearing contains much noise in actual work environment. Whether the model proposed in this paper still has a good diagnostic effect in noisy environment? Any signal noise reduction techniques can be considered?

3. When designing the basic model Deep Neural Network, the number of neurons in each layer is 600/200/101/50/30/30/9. What is the reason for designing this structure? 

Author Response

Please see the attachment.

Author Response File: Author Response.pdf

Reviewer 3 Report

The presented paper has a detailed introduction describing the current state of the problem. In the following section, the theory that is used in the solving problem is sufficiently elaborated. The experimental part is processed at a adequate level as well.

I recommend the submitted paper for publication after minor revisions.

- modify the formatting of symbols from equations in the text,

- replace Fig. 5 with a higher quality one and add a description to the figure,

- in Fig. 6 add a legend and description of individual axes,

- in Fig. 9 adjust the readability of the axis,

- edit reference [35], which is non-functional.

Author Response

Please see the attachment。

Author Response File: Author Response.pdf