Three-Phase Inverter Fault Diagnosis Based on an Improved Deep Residual Network
Round 1
Reviewer 1 Report
In this study, open-circuit fault diagnosis of three-phase inverters is achieved through a series of operations such as data enhancement, denoising, and grey-scale image transformation and deep residual networks. However, the article does not have enough space for the experimental operation part, and the experimental conclusions drawn are not convincing enough. The following comments are made for the content of this paper for reference:
Opinion 1: The length percentage of the article needs to be adjusted; the existing includes too many sections on the principles of the model while the description related to the experiment is not complete enough. The subheadings in Chapter 3 do not match the structure of the article, and a duplicate appears in line 154. “The fault diagnosis method proposed in this paper consists of five stages.”
Opinion 2: Regarding the denoising of the data, from the example in Fig. 5, it seems that the signal graph with noise can also distinguish different failure modes well, please explain the necessity of the denoising operation; meanwhile, the signal carries noise with a relatively regular pattern, is it possible to get a better noise reduction effect by using the general filtering method? Please explain the advantage or necessity of using the WPD method; the data in the experiment seems to be enhanced before denoising, please explain why it is operated in this order.
Opinion 3: Section 4.1 mentions that the loss function values are monitored to adaptively adjust the learning rate when the CVAE model is enhancing the data, but the correlation between the timing of the adjustment of the learning rate and the change in the loss function values in Fig. 13 does not show a correlation between them, so please check the data, as a clearer way of presenting the effect of the use of the dynamic learning rate may be needed.
Opinion 4: It is suggested that the signal plots generated by the CVAE model are given in the paper to illustrate the effect of data enhancement, and the current visual clustering results of t-SNE are too abstract. In addition, the results in Fig. 14 show that the difference between different failure modes is very obvious, if the 2 components extracted by the t-SNE method are used as features for the classification model, can good prediction results be achieved as well, and at the same time, due to the fact that this clustering-based method is much simpler to compute, it may have better computational efficiency?
Opinion 5: The process of data experimentation lacks the description of some key parameters, such as the size of the convolutional kernel in the classification model and the parameters of the BPNN network used for comparison. Especially the process of data set division, it seems that all the 550 samples are currently used as both training and testing sets, which gives unconvincing accuracy results, and it is recommended to proportionally distribute the data set and conduct K-fold multiple experiments before comparison.
Need to be improved
Author Response
Dear Reviewer,
Thank you for your valuable comments and guidance on my manuscript. I have made targeted revisions and provided answers to the questions you raised, the details of which can be found in the attached file.
Your critiques and corrections have been immensely helpful, and I am genuinely thankful for them. Should you have any further questions or require additional information, please do not hesitate to ask.
Thank you once again for your time and assistance.
Best regards,
Yu Ji
Author Response File: 
Author Response.pdf
Reviewer 2 Report
The authors of this study address low accuracy in fault diagnosis methods for three-phase inverters, noise interference, and fault sample limitations. To improve samples, they use Conditional Variational Auto-Encoder and Wavelet Packet Decomposition denoising. An enhanced deep residual network (SE-ResNet18) fault diagnosis model with a channel attention mechanism is used to normalize, pre-process, and train the dataset. According to the authors, the augmented fault samples increase diagnosis accuracy by achieving higher accuracy with fewer iterations and faster convergence, demonstrating its efficacy in correctly diagnosing inverter open-circuit faults across different sample scenarios.
My observations are as follows:
1. The study's limitations are not discussed.
2. The Research Founding section could be improved. A lot has already been accomplished previously. Please try to include/discuss at least five additional recent works.
3. It is advisable to review your English usage for clarity one more time.
4. In the conclusion section, please refrain from using bullets.
It is advisable to review English usage one more time for the purpose of clarity.
Author Response
Dear Reviewer,
Thank you for your valuable comments and guidance on my manuscript. I have made targeted revisions and provided answers to the questions you raised, the details of which can be found in the attached file.
Your critiques and corrections have been immensely helpful, and I am genuinely thankful for them. Should you have any further questions or require additional information, please do not hesitate to ask.
Thank you once again for your time and assistance.
Best regards,
Yu Ji
Author Response File: Author Response.pdf
Reviewer 3 Report
This paper used Wavelet Packet Decomposition (WPD) denoising and 12 Conditional Variational Auto-Encoder (CVAE) are used for sample enhancement based on the existing faulty samples. Results show that the augmented fault samples improve the diagnosis accuracy com-16 pared to the original samples. There are some suggestions are as follows:
1. Abstract and body are always considered as two separate parts. Please give the full name when
CVAE and WPD were first used showed in Introduction.
2. Authors should adjust the whole paragraphs to fit the requirement of a research paper. The suggested organization of paper should have been shown as follows: Introduction, Deep residual network, The proposed method, Experimental verifications and Conclusions.
3. As we know, the ResNet model is an updated version of the ConvNet model and is different from traditional deep learning. So, authors should tell reader what’s main differences between them?
4. Authors used CVAE to enhance the fault sample. However, they didn’t compare with other methods, like GAN?
5. It’s better to use a flowchart to present the preprocess procedures.
Minor revisions
Author Response
Dear Reviewer,
Thank you for your valuable comments and guidance on my manuscript. I have made targeted revisions and provided answers to the questions you raised, the details of which can be found in the attached file.
Your critiques and corrections have been immensely helpful, and I am genuinely thankful for them. Should you have any further questions or require additional information, please do not hesitate to ask.
Thank you once again for your time and assistance.
Best regards,
Yu Ji
Author Response File: Author Response.pdf
