Fault Detection of Bearing by Resnet Classifier with Model-Based Data Augmentation
Round 1
Reviewer 1 Report
It is a well-prepared paper.
Although the research goal is clear, the discussion is acceptable and the data processing method is new, the novelty of the approach is questionable.
It is publishable in the current form.
Author Response
Response to Reviewer 1 Comments
Point 1: It is a well-prepared paper. Although the research goal is clear, the discussion is acceptable and the data processing method is new, the novelty of the approach is questionable. It is publishable in the current form.
Response 1: Thank you very much for your approval. This paper is devoted to develop a novel fault detection scheme of bearings to deal with the fact that the available data rarely cover all the operating modes of a system, which integrates model-based data augmentation and deep learning method. For our purpose, a dynamic model of roller bearing is established to reflect the correspondence between bearing states and vibration signals and further achieve the data augmentation. To evaluate the performance of the simulation results for data augmentation, an error index is proposed. In addition, the envelop signals are instead of the original signals in the training process to reduce the gap between the simulated data and the real data. To the best of our knowledge, such a dynamic model of roller bearing and the envelop data processing are first applied in the data augmentation for bearing fault detection, which constitutes the main points of this article. Deep residual network (Resnet) is a deep leaning method with extremely deep architecture, which is a well-developed method with outstanding performance on accuracy and convergence. As the Resnet shows high enough accuracy of classification (100%) and achieves the effect we expected in our case. Therefore, this paper doesn’t focus on the further improvement of the network structure of the Resnet method, which will be one of the next directions of our research.
Author Response File: 
Author Response.pdf
Reviewer 2 Report
This paper deals with a hybrid method for roiling bearing fault diagnosis that combinesa dynamic model used to complete available data-sets by simulated data. Then, the augmented data-set is used for training fault classifier by deep learning algorithm. Moreover, the envelop signals are instead of the original signals in the training process to reduce the gap between the simulated data and the real data. Finally, the operation states of the roller bearings can be identified by the trained fault classifier by inputting the vibration signals to be classified.
The paper is well written and easy to read and understand, the problem addressed is real, it concerns the fact that the available data rarely cover all the operating modes of a system, in particular degraded or faulty operations.
The paper can be improved by the following points:
1) The proposed dynamic model will interest many readers, a nomenclature table that recalls the meaning of each parameter of the model will improve the understanding of the dynamic model.
2) When a dynamic model is used to simulate faulty or degraded functioning, the parameters of the model are varied, in this case the identification of the initial values of the parameters (representing normal functioning) becomes important. The authors must in my opinion give the method used for the parametric identification. Some identification methods converge towards local optimums which give erroneous parameter values.
3) The introduction of the paper can be enhanced by review papers on diagnostic methods in the general case as for example: A survey of fault diagnosis and fault-tolerant techniques; part i: fault diagnosis with model-based and signal-based approaches, IEEE Trans. Ind. Electron. 62 (6) (2015) 3757e3767. and other methods that can be used to augment simulation databases such as New Approach for Failure Prognosis Using a Bond Graph, Gaussian Mixture Model and Similarity Techniques".Processes 2022,10, 435. Or method using continuous Markovian models like wiener process.
Author Response
Response to Reviewer 2 Comments
Point 1: The proposed dynamic model will interest many readers, a nomenclature table that recalls the meaning of each parameter of the model will improve the understanding of the dynamic model.
Response 1: Thank you for the nice suggestion. A nomenclature table is added in the revised paper for readability and understanding of the dynamic model. The nomenclature table is given as follows.
Point 2: When a dynamic model is used to simulate faulty or degraded functioning, the parameters of the model are varied, in this case the identification of the initial values of the parameters (representing normal functioning) becomes important. The authors must in my opinion give the method used for the parametric identification. Some identification methods converge towards local optimums which give erroneous parameter values.
Response 2: Thank you for pointing this out. The process of parameter identification has been added in the revised manuscript.
Point 3: The introduction of the paper can be enhanced by review papers on diagnostic methods in the general case as for example: A survey of fault diagnosis and fault-tolerant techniques; part i: fault diagnosis with model-based and signal-based approaches, IEEE Trans. Ind. Electron. 62 (6) (2015) 3757e3767. and other methods that can be used to augment simulation databases such as New Approach for Failure Prognosis Using a Bond Graph, Gaussian Mixture Model and Similarity Techniques". Processes 2022,10, 435. Or method using continuous Markovian models like wiener process.
Response 3: Thank you very much. Some papers related to diagnostic methods in the general case and other methods have been referenced in the revised manuscript to enhanced the introduction of the paper. Please see the revised manuscript for the details.
Author Response File: Author Response.pdf
Reviewer 3 Report
All the remarks please find in attached file
Comments for author File: 
Comments.doc
Author Response
Response to Reviewer 3 Comments
Point 1: Some editorial corrections should be done e.g.
Response 1: Thank you for examining the manuscript so carefully and the nice suggestions. The corrections have been done in the revised manuscript.
Point 2: In introduction part, the works of system - based fault identification (e.g. https://0-doi-org.brum.beds.ac.uk/10.4271/03-15-04-0028) and CNN applications (e.g. https://0-doi-org.brum.beds.ac.uk/10.3390/s20144017) should be mentioned.
Response 2: Thank you for pointing this out. The papers related to system-based fault identification and other methods such as CNN have been referenced in the revised manuscript to enhanced the introduction of the paper. Please see the revised manuscript for the details.
Point 3: The formulas should be numbered.
Response 3: Thank you very much. The formulas in the revised paper have been numbered.
Point 4: Some quality measures of the model used to data augmentation should be presented in addition to the graphical comparison between the model output and the real data.
Response 4: Thank you very much for the nice suggestion.
An error index is used to evaluate the distance between the simulation results and experimental measured results. To consider the influence of the wave shift, this index is defined in frequency domain as Eq. (1).
The error index is used to evaluate the performance of the simulation results for data augmentation. The error index for the original data is 0.9214, while the error index for the envelop data is 0.4622. The gap between the real and simulation results of the original data is much larger than that of the envelop data. That is the reason why we use the envelop data as the training dataset.
Point 5: As performance measures accuracy was used. But in classification issues very often complete confusion matrix is considered. It should be explained why only accuracy is significant in case of this research.
Response 5: Thank you for the question. In general, the confusion matrix is important for the classification. In our case, the accuracy of fault detection is 100%, which results that the confusion matrix is a unit diagonal matrix, as shown in Fig. 1. For this reason, the complete confusion matrix is not given in manuscript.
Author Response File: Author Response.pdf
